[TTS] [黑客松]Add JETS (#3109)

examples/csmsc/jets/README.md

@@ -0,0 +1,98 @@
+# JETS with CSMSC
+This example contains code used to train a [JETS](https://arxiv.org/abs/2203.16852v1) model with [Chinese Standard Mandarin Speech Copus](https://www.data-baker.com/open_source.html).
+
+## Dataset
+### Download and Extract
+Download CSMSC from it's [Official Website](https://test.data-baker.com/data/index/source).
+
+### Get MFA Result and Extract
+We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get phonemes and durations for JETS.
+You can download from here [baker_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/BZNSYP/with_tone/baker_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) of our repo.
+
+## Get Started
+Assume the path to the dataset is `~/datasets/BZNSYP`.
+Assume the path to the MFA result of CSMSC is `./baker_alignment_tone`.
+Run the command below to
+1. **source path**.
+2. preprocess the dataset.
+3. train the model.
+4. synthesize wavs.
+    - synthesize waveform from `metadata.jsonl`.
+    - synthesize waveform from a text file.
+
+```bash
+./run.sh
+```
+You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, running the following command will only preprocess the dataset.
+```bash
+./run.sh --stage 0 --stop-stage 0
+```
+### Data Preprocessing
+```bash
+./local/preprocess.sh ${conf_path}
+```
+When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
+
+```text
+dump
+├── dev
+│   ├── norm
+│   └── raw
+├── phone_id_map.txt
+├── speaker_id_map.txt
+├── test
+│   ├── norm
+│   └── raw
+└── train
+    ├── feats_stats.npy
+    ├── norm
+    └── raw
+```
+The dataset is split into 3 parts, namely `train`, `dev`, and` test`, each of which contains a `norm` and `raw` subfolder. The raw folder contains wave、mel spectrogram、speech、pitch and energy features of each utterance, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/feats_stats.npy`.
+
+Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, the path of feats, feats_lengths, the path of pitch features, the path of energy features, the path of raw waves, speaker, and the id of each utterance.
+
+### Model Training
+```bash
+CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
+```
+`./local/train.sh` calls `${BIN_DIR}/train.py`.
+Here's the complete help message.
+```text
+usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
+                [--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
+                [--ngpu NGPU] [--phones-dict PHONES_DICT]
+
+Train a JETS model.
+
+optional arguments:
+  -h, --help            show this help message and exit
+  --config CONFIG       config file to overwrite default config.
+  --train-metadata TRAIN_METADATA
+                        training data.
+  --dev-metadata DEV_METADATA
+                        dev data.
+  --output-dir OUTPUT_DIR
+                        output dir.
+  --ngpu NGPU           if ngpu == 0, use cpu.
+  --phones-dict PHONES_DICT
+                        phone vocabulary file.
+```
+1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
+2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
+3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are saved in `checkpoints/` inside this directory.
+4. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
+5. `--phones-dict` is the path of the phone vocabulary file.
+
+### Synthesizing
+
+`./local/synthesize.sh` calls `${BIN_DIR}/synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
+
+```bash
+CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
+```
+
+`./local/synthesize_e2e.sh` calls `${BIN_DIR}/synthesize_e2e.py`, which can synthesize waveform from text file.
+```bash
+CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name}
+```

examples/csmsc/jets/conf/default.yaml

@@ -0,0 +1,224 @@
+# This configuration tested on 4 GPUs (V100) with 32GB GPU
+# memory. It takes around 2 weeks to finish the training
+# but 100k iters model should generate reasonable results.
+###########################################################
+#                FEATURE EXTRACTION SETTING               #
+###########################################################
+
+n_mels: 80
+fs: 22050         # sr
+n_fft: 1024        # FFT size (samples).
+n_shift: 256       # Hop size (samples). 12.5ms
+win_length: null   # Window length (samples). 50ms
+                   # If set to null, it will be the same as fft_size.
+window: "hann"     # Window function.
+fmin: 0            # minimum frequency for Mel basis
+fmax: null         # maximum frequency for Mel basis
+f0min: 80          # Minimum f0 for pitch extraction.
+f0max: 400         # Maximum f0 for pitch extraction.
+
+
+##########################################################
+#                  TTS MODEL SETTING                     #
+##########################################################
+model:
+    # generator related
+    generator_type: jets_generator
+    generator_params:
+        adim: 256         # attention dimension
+        aheads: 2         # number of attention heads
+        elayers: 4        # number of encoder layers
+        eunits: 1024      # number of encoder ff units
+        dlayers: 4        # number of decoder layers
+        dunits: 1024      # number of decoder ff units
+        positionwise_layer_type: conv1d   # type of position-wise layer
+        positionwise_conv_kernel_size: 3  # kernel size of position wise conv layer
+        duration_predictor_layers: 2      # number of layers of duration predictor
+        duration_predictor_chans: 256     # number of channels of duration predictor
+        duration_predictor_kernel_size: 3 # filter size of duration predictor
+        use_masking: True                 # whether to apply masking for padded part in loss calculation
+        encoder_normalize_before: True    # whether to perform layer normalization before the input
+        decoder_normalize_before: True    # whether to perform layer normalization before the input
+        encoder_type: transformer           # encoder type
+        decoder_type: transformer           # decoder type
+        conformer_rel_pos_type: latest               # relative positional encoding type
+        conformer_pos_enc_layer_type: rel_pos        # conformer positional encoding type
+        conformer_self_attn_layer_type: rel_selfattn # conformer self-attention type
+        conformer_activation_type: swish             # conformer activation type
+        use_macaron_style_in_conformer: true         # whether to use macaron style in conformer
+        use_cnn_in_conformer: true                   # whether to use CNN in conformer
+        conformer_enc_kernel_size: 7                 # kernel size in CNN module of conformer-based encoder
+        conformer_dec_kernel_size: 31                # kernel size in CNN module of conformer-based decoder
+        init_type: xavier_uniform                    # initialization type
+        init_enc_alpha: 1.0                          # initial value of alpha for encoder
+        init_dec_alpha: 1.0                          # initial value of alpha for decoder
+        transformer_enc_dropout_rate: 0.2            # dropout rate for transformer encoder layer
+        transformer_enc_positional_dropout_rate: 0.2 # dropout rate for transformer encoder positional encoding
+        transformer_enc_attn_dropout_rate: 0.2       # dropout rate for transformer encoder attention layer
+        transformer_dec_dropout_rate: 0.2            # dropout rate for transformer decoder layer
+        transformer_dec_positional_dropout_rate: 0.2 # dropout rate for transformer decoder positional encoding
+        transformer_dec_attn_dropout_rate: 0.2       # dropout rate for transformer decoder attention layer
+        pitch_predictor_layers: 5                    # number of conv layers in pitch predictor
+        pitch_predictor_chans: 256                   # number of channels of conv layers in pitch predictor
+        pitch_predictor_kernel_size: 5               # kernel size of conv leyers in pitch predictor
+        pitch_predictor_dropout: 0.5                 # dropout rate in pitch predictor
+        pitch_embed_kernel_size: 1                   # kernel size of conv embedding layer for pitch
+        pitch_embed_dropout: 0.0                     # dropout rate after conv embedding layer for pitch
+        stop_gradient_from_pitch_predictor: true     # whether to stop the gradient from pitch predictor to encoder
+        energy_predictor_layers: 2                   # number of conv layers in energy predictor
+        energy_predictor_chans: 256                  # number of channels of conv layers in energy predictor
+        energy_predictor_kernel_size: 3              # kernel size of conv leyers in energy predictor
+        energy_predictor_dropout: 0.5                # dropout rate in energy predictor
+        energy_embed_kernel_size: 1                  # kernel size of conv embedding layer for energy
+        energy_embed_dropout: 0.0                    # dropout rate after conv embedding layer for energy
+        stop_gradient_from_energy_predictor: false   # whether to stop the gradient from energy predictor to encoder
+        generator_out_channels: 1
+        generator_channels: 512
+        generator_global_channels: -1
+        generator_kernel_size: 7
+        generator_upsample_scales: [8, 8, 2, 2]
+        generator_upsample_kernel_sizes: [16, 16, 4, 4]
+        generator_resblock_kernel_sizes: [3, 7, 11]
+        generator_resblock_dilations: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+        generator_use_additional_convs: true
+        generator_bias: true
+        generator_nonlinear_activation: "leakyrelu"
+        generator_nonlinear_activation_params:
+            negative_slope: 0.1
+        generator_use_weight_norm: true
+        segment_size: 64              # segment size for random windowed discriminator
+
+    # discriminator related
+    discriminator_type: hifigan_multi_scale_multi_period_discriminator
+    discriminator_params:
+        scales: 1
+        scale_downsample_pooling: "AvgPool1D"
+        scale_downsample_pooling_params:
+            kernel_size: 4
+            stride: 2
+            padding: 2
+        scale_discriminator_params:
+            in_channels: 1
+            out_channels: 1
+            kernel_sizes: [15, 41, 5, 3]
+            channels: 128
+            max_downsample_channels: 1024
+            max_groups: 16
+            bias: True
+            downsample_scales: [2, 2, 4, 4, 1]
+            nonlinear_activation: "leakyrelu"
+            nonlinear_activation_params:
+                negative_slope: 0.1
+            use_weight_norm: True
+            use_spectral_norm: False
+        follow_official_norm: False
+        periods: [2, 3, 5, 7, 11]
+        period_discriminator_params:
+            in_channels: 1
+            out_channels: 1
+            kernel_sizes: [5, 3]
+            channels: 32
+            downsample_scales: [3, 3, 3, 3, 1]
+            max_downsample_channels: 1024
+            bias: True
+            nonlinear_activation: "leakyrelu"
+            nonlinear_activation_params:
+                negative_slope: 0.1
+            use_weight_norm: True
+            use_spectral_norm: False
+    # others
+    sampling_rate: 22050          # needed in the inference for saving wav
+    cache_generator_outputs: True # whether to cache generator outputs in the training
+use_alignment_module: False       # whether to use alignment module
+      
+###########################################################
+#                        LOSS SETTING                     #
+###########################################################
+# loss function related
+generator_adv_loss_params:
+    average_by_discriminators: False # whether to average loss value by #discriminators
+    loss_type: mse                   # loss type, "mse" or "hinge"
+discriminator_adv_loss_params:
+    average_by_discriminators: False # whether to average loss value by #discriminators
+    loss_type: mse                   # loss type, "mse" or "hinge"
+feat_match_loss_params:
+    average_by_discriminators: False # whether to average loss value by #discriminators
+    average_by_layers: False         # whether to average loss value by #layers of each discriminator
+    include_final_outputs: True      # whether to include final outputs for loss calculation
+mel_loss_params:
+    fs: 22050          # must be the same as the training data
+    fft_size: 1024        # fft points
+    hop_size: 256    # hop size
+    win_length: null   # window length
+    window: hann       # window type
+    num_mels: 80         # number of Mel basis
+    fmin: 0            # minimum frequency for Mel basis
+    fmax: null         # maximum frequency for Mel basis
+    log_base: null     # null represent natural log
+
+###########################################################
+#               ADVERSARIAL LOSS SETTING                  #
+###########################################################
+lambda_adv: 1.0        # loss scaling coefficient for adversarial loss
+lambda_mel: 45.0       # loss scaling coefficient for Mel loss
+lambda_feat_match: 2.0 # loss scaling coefficient for feat match loss
+lambda_var: 1.0        # loss scaling coefficient for duration loss
+lambda_align: 2.0         # loss scaling coefficient for KL divergence loss
+# others
+sampling_rate: 22050          # needed in the inference for saving wav
+cache_generator_outputs: True # whether to cache generator outputs in the training
+
+
+# extra module for additional inputs
+pitch_extract: dio           # pitch extractor type
+pitch_extract_conf:
+    reduction_factor: 1
+    use_token_averaged_f0: false
+pitch_normalize: global_mvn  # normalizer for the pitch feature
+energy_extract: energy       # energy extractor type
+energy_extract_conf:
+    reduction_factor: 1
+    use_token_averaged_energy: false
+energy_normalize: global_mvn # normalizer for the energy feature
+
+
+###########################################################
+#                  DATA LOADER SETTING                    #
+###########################################################
+batch_size: 32              # Batch size.
+num_workers: 4              # Number of workers in DataLoader.
+
+##########################################################
+#            OPTIMIZER & SCHEDULER SETTING               #
+##########################################################
+# optimizer setting for generator
+generator_optimizer_params:
+    beta1: 0.8
+    beta2: 0.99
+    epsilon: 1.0e-9
+    weight_decay: 0.0
+generator_scheduler: exponential_decay
+generator_scheduler_params:
+    learning_rate: 2.0e-4
+    gamma: 0.999875                   
+
+# optimizer setting for discriminator
+discriminator_optimizer_params:
+    beta1: 0.8
+    beta2: 0.99
+    epsilon: 1.0e-9
+    weight_decay: 0.0
+discriminator_scheduler: exponential_decay
+discriminator_scheduler_params:
+    learning_rate: 2.0e-4          
+    gamma: 0.999875
+generator_first: True # whether to start updating generator first
+
+##########################################################
+#                OTHER TRAINING SETTING                  #
+##########################################################
+num_snapshots: 10            # max number of snapshots to keep while training
+train_max_steps: 350000      # Number of training steps. == total_iters / ngpus, total_iters = 1000000
+save_interval_steps: 1000    # Interval steps to save checkpoint.
+eval_interval_steps: 250     # Interval steps to evaluate the network.
+seed: 777                    # random seed number

examples/csmsc/jets/local/inference.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+
+train_output_path=$1
+
+stage=0
+stop_stage=0
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    python3 ${BIN_DIR}/inference.py \
+        --inference_dir=${train_output_path}/inference \
+        --am=jets_csmsc \
+        --text=${BIN_DIR}/../sentences.txt \
+        --output_dir=${train_output_path}/pd_infer_out \
+        --phones_dict=dump/phone_id_map.txt
+fi

examples/csmsc/jets/local/preprocess.sh

@@ -0,0 +1,77 @@
+#!/bin/bash
+set -e
+stage=0
+stop_stage=100
+
+config_path=$1
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # get durations from MFA's result
+    echo "Generate durations.txt from MFA results ..."
+    python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
+        --inputdir=./baker_alignment_tone \
+        --output=durations.txt \
+        --config=${config_path}
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # extract features
+    echo "Extract features ..."
+    python3 ${BIN_DIR}/preprocess.py \
+        --dataset=baker \
+        --rootdir=~/datasets/BZNSYP/ \
+        --dumpdir=dump \
+        --dur-file=durations.txt \
+        --config=${config_path} \
+        --num-cpu=20 \
+        --cut-sil=True \
+        --token_average=True
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    # get features' stats(mean and std)
+    echo "Get features' stats ..."
+    python3 ${MAIN_ROOT}/utils/compute_statistics.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --field-name="feats"
+
+    python3 ${MAIN_ROOT}/utils/compute_statistics.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --field-name="pitch"
+
+    python3 ${MAIN_ROOT}/utils/compute_statistics.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --field-name="energy"
+
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    # normalize and covert phone/speaker to id, dev and test should use train's stats
+    echo "Normalize ..."
+    python3 ${BIN_DIR}/normalize.py \
+        --metadata=dump/train/raw/metadata.jsonl \
+        --dumpdir=dump/train/norm \
+        --feats-stats=dump/train/feats_stats.npy \
+        --pitch-stats=dump/train/pitch_stats.npy \
+        --energy-stats=dump/train/energy_stats.npy \
+        --phones-dict=dump/phone_id_map.txt \
+        --speaker-dict=dump/speaker_id_map.txt
+
+    python3 ${BIN_DIR}/normalize.py \
+        --metadata=dump/dev/raw/metadata.jsonl \
+        --dumpdir=dump/dev/norm \
+        --feats-stats=dump/train/feats_stats.npy \
+        --pitch-stats=dump/train/pitch_stats.npy \
+        --energy-stats=dump/train/energy_stats.npy \
+        --phones-dict=dump/phone_id_map.txt \
+        --speaker-dict=dump/speaker_id_map.txt
+
+    python3 ${BIN_DIR}/normalize.py \
+        --metadata=dump/test/raw/metadata.jsonl \
+        --dumpdir=dump/test/norm \
+        --feats-stats=dump/train/feats_stats.npy \
+        --pitch-stats=dump/train/pitch_stats.npy \
+        --energy-stats=dump/train/energy_stats.npy \
+        --phones-dict=dump/phone_id_map.txt \
+        --speaker-dict=dump/speaker_id_map.txt
+fi

examples/csmsc/jets/local/synthesize.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+config_path=$1
+train_output_path=$2
+ckpt_name=$3
+stage=0
+stop_stage=0
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    FLAGS_allocator_strategy=naive_best_fit \
+    FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+    python3 ${BIN_DIR}/synthesize.py \
+        --config=${config_path} \
+        --ckpt=${train_output_path}/checkpoints/${ckpt_name} \
+        --phones_dict=dump/phone_id_map.txt \
+        --test_metadata=dump/test/norm/metadata.jsonl \
+        --output_dir=${train_output_path}/test
+fi

examples/csmsc/jets/local/synthesize_e2e.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+config_path=$1
+train_output_path=$2
+ckpt_name=$3
+
+stage=0
+stop_stage=0
+
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    FLAGS_allocator_strategy=naive_best_fit \
+    FLAGS_fraction_of_gpu_memory_to_use=0.01 \
+    python3 ${BIN_DIR}/synthesize_e2e.py \
+        --am=jets_csmsc \
+        --config=${config_path} \
+        --ckpt=${train_output_path}/checkpoints/${ckpt_name} \
+        --phones_dict=dump/phone_id_map.txt \
+        --output_dir=${train_output_path}/test_e2e \
+        --text=${BIN_DIR}/../sentences.txt \
+        --inference_dir=${train_output_path}/inference
+fi

examples/csmsc/jets/local/train.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+config_path=$1
+train_output_path=$2
+
+python3 ${BIN_DIR}/train.py \
+    --train-metadata=dump/train/norm/metadata.jsonl \
+    --dev-metadata=dump/dev/norm/metadata.jsonl \
+    --config=${config_path} \
+    --output-dir=${train_output_path} \
+    --ngpu=1 \
+    --phones-dict=dump/phone_id_map.txt

examples/csmsc/jets/path.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+export MAIN_ROOT=`realpath ${PWD}/../../../`
+
+export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
+export LC_ALL=C
+
+export PYTHONDONTWRITEBYTECODE=1
+# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
+export PYTHONIOENCODING=UTF-8
+export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
+
+MODEL=jets
+export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}

examples/csmsc/jets/run.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+
+set -e
+source path.sh
+
+gpus=0
+stage=0
+stop_stage=100
+
+conf_path=conf/default.yaml
+train_output_path=exp/default
+ckpt_name=snapshot_iter_150000.pdz
+
+# with the following command, you can choose the stage range you want to run
+# such as `./run.sh --stage 0 --stop-stage 0`
+# this can not be mixed use with `$1`, `$2` ...
+source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    # prepare data
+    ./local/preprocess.sh ${conf_path}|| exit -1
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    # train model, all `ckpt` under `train_output_path/checkpoints/` dir
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+    # synthesize_e2e
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize_e2e.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
+fi
+
+if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+    CUDA_VISIBLE_DEVICES=${gpus} ./local/inference.sh ${train_output_path} || exit -1
+fi
+

paddlespeech/t2s/datasets/am_batch_fn.py

@@ -669,6 +669,142 @@ def vits_multi_spk_batch_fn(examples):
     return batch
 
 
+def jets_single_spk_batch_fn(examples):
+    """
+    Returns:
+        Dict[str, Any]:
+            - text (Tensor): Text index tensor (B, T_text).
+            - text_lengths (Tensor): Text length tensor (B,).
+            - feats (Tensor): Feature tensor (B, T_feats, aux_channels).
+            - feats_lengths (Tensor): Feature length tensor (B,).
+            - durations (Tensor): Feature tensor (B, T_text,).
+            - durations_lengths (Tensor): Durations length tensor (B,).
+            - pitch (Tensor): Feature tensor (B, pitch_length,).
+            - energy (Tensor): Feature tensor (B, energy_length,).
+            - speech (Tensor): Speech waveform tensor (B, T_wav).
+
+    """
+    # fields = ["text", "text_lengths", "feats", "feats_lengths", "durations", "pitch", "energy", "speech"]
+    text = [np.array(item["text"], dtype=np.int64) for item in examples]
+    feats = [np.array(item["feats"], dtype=np.float32) for item in examples]
+    durations = [
+        np.array(item["durations"], dtype=np.int64) for item in examples
+    ]
+    pitch = [np.array(item["pitch"], dtype=np.float32) for item in examples]
+    energy = [np.array(item["energy"], dtype=np.float32) for item in examples]
+    speech = [np.array(item["wave"], dtype=np.float32) for item in examples]
+
+    text_lengths = [
+        np.array(item["text_lengths"], dtype=np.int64) for item in examples
+    ]
+    feats_lengths = [
+        np.array(item["feats_lengths"], dtype=np.int64) for item in examples
+    ]
+
+    text = batch_sequences(text)
+    feats = batch_sequences(feats)
+    durations = batch_sequences(durations)
+    pitch = batch_sequences(pitch)
+    energy = batch_sequences(energy)
+    speech = batch_sequences(speech)
+
+    # convert each batch to paddle.Tensor
+    text = paddle.to_tensor(text)
+    feats = paddle.to_tensor(feats)
+    durations = paddle.to_tensor(durations)
+    pitch = paddle.to_tensor(pitch)
+    energy = paddle.to_tensor(energy)
+    text_lengths = paddle.to_tensor(text_lengths)
+    feats_lengths = paddle.to_tensor(feats_lengths)
+
+    batch = {
+        "text": text,
+        "text_lengths": text_lengths,
+        "feats": feats,
+        "feats_lengths": feats_lengths,
+        "durations": durations,
+        "durations_lengths": text_lengths,
+        "pitch": pitch,
+        "energy": energy,
+        "speech": speech,
+    }
+    return batch
+
+
+def jets_multi_spk_batch_fn(examples):
+    """
+    Returns:
+        Dict[str, Any]:
+            - text (Tensor): Text index tensor (B, T_text).
+            - text_lengths (Tensor): Text length tensor (B,).
+            - feats (Tensor): Feature tensor (B, T_feats, aux_channels).
+            - feats_lengths (Tensor): Feature length tensor (B,).
+            - durations (Tensor): Feature tensor (B, T_text,).
+            - durations_lengths (Tensor): Durations length tensor (B,).
+            - pitch (Tensor): Feature tensor (B, pitch_length,).
+            - energy (Tensor): Feature tensor (B, energy_length,).
+            - speech (Tensor): Speech waveform tensor (B, T_wav).
+            - spk_id (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            - spk_emb (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+    """
+    # fields = ["text", "text_lengths", "feats", "feats_lengths", "durations", "pitch", "energy", "speech", "spk_id"/"spk_emb"]
+    text = [np.array(item["text"], dtype=np.int64) for item in examples]
+    feats = [np.array(item["feats"], dtype=np.float32) for item in examples]
+    durations = [
+        np.array(item["durations"], dtype=np.int64) for item in examples
+    ]
+    pitch = [np.array(item["pitch"], dtype=np.float32) for item in examples]
+    energy = [np.array(item["energy"], dtype=np.float32) for item in examples]
+    speech = [np.array(item["wave"], dtype=np.float32) for item in examples]
+    text_lengths = [
+        np.array(item["text_lengths"], dtype=np.int64) for item in examples
+    ]
+    feats_lengths = [
+        np.array(item["feats_lengths"], dtype=np.int64) for item in examples
+    ]
+
+    text = batch_sequences(text)
+    feats = batch_sequences(feats)
+    durations = batch_sequences(durations)
+    pitch = batch_sequences(pitch)
+    energy = batch_sequences(energy)
+    speech = batch_sequences(speech)
+
+    # convert each batch to paddle.Tensor
+    text = paddle.to_tensor(text)
+    feats = paddle.to_tensor(feats)
+    durations = paddle.to_tensor(durations)
+    pitch = paddle.to_tensor(pitch)
+    energy = paddle.to_tensor(energy)
+    text_lengths = paddle.to_tensor(text_lengths)
+    feats_lengths = paddle.to_tensor(feats_lengths)
+
+    batch = {
+        "text": text,
+        "text_lengths": text_lengths,
+        "feats": feats,
+        "feats_lengths": feats_lengths,
+        "durations": durations,
+        "durations_lengths": text_lengths,
+        "pitch": pitch,
+        "energy": energy,
+        "speech": speech,
+    }
+    # spk_emb has a higher priority than spk_id
+    if "spk_emb" in examples[0]:
+        spk_emb = [
+            np.array(item["spk_emb"], dtype=np.float32) for item in examples
+        ]
+        spk_emb = batch_sequences(spk_emb)
+        spk_emb = paddle.to_tensor(spk_emb)
+        batch["spk_emb"] = spk_emb
+    elif "spk_id" in examples[0]:
+        spk_id = [np.array(item["spk_id"], dtype=np.int64) for item in examples]
+        spk_id = paddle.to_tensor(spk_id)
+        batch["spk_id"] = spk_id
+    return batch
+
+
 # 因为要传参数，所以需要额外构建
 def build_starganv2_vc_collate_fn(latent_dim: int=16, max_mel_length: int=192):
 

paddlespeech/t2s/exps/jets/__init__.py

@@ -0,0 +1,13 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

paddlespeech/t2s/exps/jets/inference.py

@@ -0,0 +1,172 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+from pathlib import Path
+
+import paddle
+import soundfile as sf
+from timer import timer
+
+from paddlespeech.t2s.exps.syn_utils import get_am_output
+from paddlespeech.t2s.exps.syn_utils import get_frontend
+from paddlespeech.t2s.exps.syn_utils import get_predictor
+from paddlespeech.t2s.exps.syn_utils import get_sentences
+from paddlespeech.t2s.utils import str2bool
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Paddle Infernce with acoustic model & vocoder.")
+    # acoustic model
+    parser.add_argument(
+        '--am',
+        type=str,
+        default='jets_csmsc',
+        choices=['jets_csmsc', 'jets_aishell3'],
+        help='Choose acoustic model type of tts task.')
+    parser.add_argument(
+        "--phones_dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--speaker_dict", type=str, default=None, help="speaker id map file.")
+    parser.add_argument(
+        '--spk_id',
+        type=int,
+        default=0,
+        help='spk id for multi speaker acoustic model')
+    # other
+    parser.add_argument(
+        '--lang',
+        type=str,
+        default='zh',
+        help='Choose model language. zh or en or mix')
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="text to synthesize, a 'utt_id sentence' pair per line")
+    parser.add_argument(
+        "--add-blank",
+        type=str2bool,
+        default=True,
+        help="whether to add blank between phones")
+    parser.add_argument(
+        "--inference_dir", type=str, help="dir to save inference models")
+    parser.add_argument("--output_dir", type=str, help="output dir")
+    # inference
+    parser.add_argument(
+        "--use_trt",
+        type=str2bool,
+        default=False,
+        help="whether to use TensorRT or not in GPU", )
+    parser.add_argument(
+        "--use_mkldnn",
+        type=str2bool,
+        default=False,
+        help="whether to use MKLDNN or not in CPU.", )
+    parser.add_argument(
+        "--precision",
+        type=str,
+        default='fp32',
+        choices=['fp32', 'fp16', 'bf16', 'int8'],
+        help="mode of running")
+    parser.add_argument(
+        "--device",
+        default="gpu",
+        choices=["gpu", "cpu"],
+        help="Device selected for inference.", )
+    parser.add_argument('--cpu_threads', type=int, default=1)
+
+    args, _ = parser.parse_known_args()
+    return args
+
+
+# only inference for models trained with csmsc now
+def main():
+    args = parse_args()
+
+    paddle.set_device(args.device)
+
+    # frontend
+    frontend = get_frontend(lang=args.lang, phones_dict=args.phones_dict)
+
+    # am_predictor
+    am_predictor = get_predictor(
+        model_dir=args.inference_dir,
+        model_file=args.am + ".pdmodel",
+        params_file=args.am + ".pdiparams",
+        device=args.device,
+        use_trt=args.use_trt,
+        use_mkldnn=args.use_mkldnn,
+        cpu_threads=args.cpu_threads,
+        precision=args.precision)
+    # model: {model_name}_{dataset}
+    am_dataset = args.am[args.am.rindex('_') + 1:]
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    sentences = get_sentences(text_file=args.text, lang=args.lang)
+
+    merge_sentences = True
+    add_blank = args.add_blank
+    # jets's fs is 22050
+    fs = 22050
+    # warmup
+    for utt_id, sentence in sentences[:3]:
+        with timer() as t:
+            wav = get_am_output(
+                input=sentence,
+                am_predictor=am_predictor,
+                am=args.am,
+                frontend=frontend,
+                lang=args.lang,
+                merge_sentences=merge_sentences,
+                speaker_dict=args.speaker_dict,
+                spk_id=args.spk_id, )
+        speed = wav.size / t.elapse
+        rtf = fs / speed
+        print(
+            f"{utt_id}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+
+    print("warm up done!")
+
+    N = 0
+    T = 0
+    for utt_id, sentence in sentences:
+        with timer() as t:
+            wav = get_am_output(
+                input=sentence,
+                am_predictor=am_predictor,
+                am=args.am,
+                frontend=frontend,
+                lang=args.lang,
+                merge_sentences=merge_sentences,
+                speaker_dict=args.speaker_dict,
+                spk_id=args.spk_id, )
+
+        N += wav.size
+        T += t.elapse
+        speed = wav.size / t.elapse
+        rtf = fs / speed
+        sf.write(output_dir / (utt_id + ".wav"), wav, samplerate=fs)
+        print(
+            f"{utt_id}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+
+        print(f"{utt_id} done!")
+    print(f"generation speed: {N / T}Hz, RTF: {fs / (N / T) }")
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/jets/normalize.py

@@ -0,0 +1,163 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Normalize feature files and dump them."""
+import argparse
+import logging
+from operator import itemgetter
+from pathlib import Path
+
+import jsonlines
+import numpy as np
+from sklearn.preprocessing import StandardScaler
+from tqdm import tqdm
+
+from paddlespeech.t2s.datasets.data_table import DataTable
+
+
+def main():
+    """Run preprocessing process."""
+    parser = argparse.ArgumentParser(
+        description="Normalize dumped raw features (See detail in parallel_wavegan/bin/normalize.py)."
+    )
+    parser.add_argument(
+        "--metadata",
+        type=str,
+        required=True,
+        help="directory including feature files to be normalized. "
+        "you need to specify either *-scp or rootdir.")
+
+    parser.add_argument(
+        "--dumpdir",
+        type=str,
+        required=True,
+        help="directory to dump normalized feature files.")
+    parser.add_argument(
+        "--feats-stats", type=str, required=True, help="feats statistics file.")
+    parser.add_argument(
+        "--pitch-stats", type=str, required=True, help="pitch statistics file.")
+    parser.add_argument(
+        "--energy-stats",
+        type=str,
+        required=True,
+        help="energy statistics file.")
+    parser.add_argument(
+        "--phones-dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--speaker-dict", type=str, default=None, help="speaker id map file.")
+
+    args = parser.parse_args()
+
+    dumpdir = Path(args.dumpdir).expanduser()
+    # use absolute path
+    dumpdir = dumpdir.resolve()
+    dumpdir.mkdir(parents=True, exist_ok=True)
+
+    # get dataset
+    with jsonlines.open(args.metadata, 'r') as reader:
+        metadata = list(reader)
+    dataset = DataTable(
+        metadata,
+        converters={
+            "feats": np.load,
+            "pitch": np.load,
+            "energy": np.load,
+            "wave": str,
+        })
+    logging.info(f"The number of files = {len(dataset)}.")
+
+    # restore scaler
+    feats_scaler = StandardScaler()
+    feats_scaler.mean_ = np.load(args.feats_stats)[0]
+    feats_scaler.scale_ = np.load(args.feats_stats)[1]
+    feats_scaler.n_features_in_ = feats_scaler.mean_.shape[0]
+
+    pitch_scaler = StandardScaler()
+    pitch_scaler.mean_ = np.load(args.pitch_stats)[0]
+    pitch_scaler.scale_ = np.load(args.pitch_stats)[1]
+    pitch_scaler.n_features_in_ = pitch_scaler.mean_.shape[0]
+
+    energy_scaler = StandardScaler()
+    energy_scaler.mean_ = np.load(args.energy_stats)[0]
+    energy_scaler.scale_ = np.load(args.energy_stats)[1]
+    energy_scaler.n_features_in_ = energy_scaler.mean_.shape[0]
+
+    vocab_phones = {}
+    with open(args.phones_dict, 'rt') as f:
+        phn_id = [line.strip().split() for line in f.readlines()]
+    for phn, id in phn_id:
+        vocab_phones[phn] = int(id)
+
+    vocab_speaker = {}
+    with open(args.speaker_dict, 'rt') as f:
+        spk_id = [line.strip().split() for line in f.readlines()]
+    for spk, id in spk_id:
+        vocab_speaker[spk] = int(id)
+
+    # process each file
+    output_metadata = []
+
+    for item in tqdm(dataset):
+        utt_id = item['utt_id']
+        feats = item['feats']
+        pitch = item['pitch']
+        energy = item['energy']
+        wave_path = item['wave']
+        # normalize
+        feats = feats_scaler.transform(feats)
+        feats_dir = dumpdir / "data_feats"
+        feats_dir.mkdir(parents=True, exist_ok=True)
+        feats_path = feats_dir / f"{utt_id}_feats.npy"
+        np.save(feats_path, feats.astype(np.float32), allow_pickle=False)
+
+        pitch = pitch_scaler.transform(pitch)
+        pitch_dir = dumpdir / "data_pitch"
+        pitch_dir.mkdir(parents=True, exist_ok=True)
+        pitch_path = pitch_dir / f"{utt_id}_pitch.npy"
+        np.save(pitch_path, pitch.astype(np.float32), allow_pickle=False)
+
+        energy = energy_scaler.transform(energy)
+        energy_dir = dumpdir / "data_energy"
+        energy_dir.mkdir(parents=True, exist_ok=True)
+        energy_path = energy_dir / f"{utt_id}_energy.npy"
+        np.save(energy_path, energy.astype(np.float32), allow_pickle=False)
+
+        phone_ids = [vocab_phones[p] for p in item['phones']]
+        spk_id = vocab_speaker[item["speaker"]]
+        record = {
+            "utt_id": item['utt_id'],
+            "spk_id": spk_id,
+            "text": phone_ids,
+            "text_lengths": item['text_lengths'],
+            "feats_lengths": item['feats_lengths'],
+            "durations": item['durations'],
+            "feats": str(feats_path),
+            "pitch": str(pitch_path),
+            "energy": str(energy_path),
+            "wave": str(wave_path),
+        }
+        # add spk_emb for voice cloning
+        if "spk_emb" in item:
+            record["spk_emb"] = str(item["spk_emb"])
+
+        output_metadata.append(record)
+    output_metadata.sort(key=itemgetter('utt_id'))
+    output_metadata_path = Path(args.dumpdir) / "metadata.jsonl"
+    with jsonlines.open(output_metadata_path, 'w') as writer:
+        for item in output_metadata:
+            writer.write(item)
+    logging.info(f"metadata dumped into {output_metadata_path}")
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/jets/preprocess.py

@@ -0,0 +1,451 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import os
+from concurrent.futures import ThreadPoolExecutor
+from operator import itemgetter
+from pathlib import Path
+from typing import Any
+from typing import Dict
+from typing import List
+
+import jsonlines
+import librosa
+import numpy as np
+import tqdm
+import yaml
+from yacs.config import CfgNode
+
+from paddlespeech.t2s.datasets.get_feats import Energy
+from paddlespeech.t2s.datasets.get_feats import LogMelFBank
+from paddlespeech.t2s.datasets.get_feats import Pitch
+from paddlespeech.t2s.datasets.preprocess_utils import compare_duration_and_mel_length
+from paddlespeech.t2s.datasets.preprocess_utils import get_input_token
+from paddlespeech.t2s.datasets.preprocess_utils import get_phn_dur
+from paddlespeech.t2s.datasets.preprocess_utils import get_spk_id_map
+from paddlespeech.t2s.datasets.preprocess_utils import merge_silence
+from paddlespeech.t2s.utils import str2bool
+
+
+def process_sentence(config: Dict[str, Any],
+                     fp: Path,
+                     sentences: Dict,
+                     output_dir: Path,
+                     mel_extractor=None,
+                     pitch_extractor=None,
+                     energy_extractor=None,
+                     cut_sil: bool=True,
+                     spk_emb_dir: Path=None,
+                     token_average: bool=True):
+    utt_id = fp.stem
+    # for vctk
+    if utt_id.endswith("_mic2"):
+        utt_id = utt_id[:-5]
+    record = None
+    if utt_id in sentences:
+        # reading, resampling may occur
+        wav, _ = librosa.load(
+            str(fp), sr=config.fs,
+            mono=False) if "canton" in str(fp) else librosa.load(
+                str(fp), sr=config.fs)
+        if len(wav.shape) == 2 and "canton" in str(fp):
+            # Remind that Cantonese datasets should be placed in ~/datasets/canton_all. Otherwise, it may cause problem.
+            wav = wav[0]
+            wav = np.ascontiguousarray(wav)
+        elif len(wav.shape) != 1:
+            return record
+        max_value = np.abs(wav).max()
+        if max_value > 1.0:
+            wav = wav / max_value
+        assert len(wav.shape) == 1, f"{utt_id} is not a mono-channel audio."
+        assert np.abs(wav).max(
+        ) <= 1.0, f"{utt_id} is seems to be different that 16 bit PCM."
+        phones = sentences[utt_id][0]
+        durations = sentences[utt_id][1]
+        speaker = sentences[utt_id][2]
+        d_cumsum = np.pad(np.array(durations).cumsum(0), (1, 0), 'constant')
+        # little imprecise than use *.TextGrid directly
+        times = librosa.frames_to_time(
+            d_cumsum, sr=config.fs, hop_length=config.n_shift)
+        if cut_sil:
+            start = 0
+            end = d_cumsum[-1]
+            if phones[0] == "sil" and len(durations) > 1:
+                start = times[1]
+                durations = durations[1:]
+                phones = phones[1:]
+            if phones[-1] == 'sil' and len(durations) > 1:
+                end = times[-2]
+                durations = durations[:-1]
+                phones = phones[:-1]
+            sentences[utt_id][0] = phones
+            sentences[utt_id][1] = durations
+            start, end = librosa.time_to_samples([start, end], sr=config.fs)
+            wav = wav[start:end]
+        # extract mel feats
+        logmel = mel_extractor.get_log_mel_fbank(wav)
+        # change duration according to mel_length
+        compare_duration_and_mel_length(sentences, utt_id, logmel)
+        # utt_id may be popped in compare_duration_and_mel_length
+        if utt_id not in sentences:
+            return None
+        phones = sentences[utt_id][0]
+        durations = sentences[utt_id][1]
+        num_frames = logmel.shape[0]
+        assert sum(durations) == num_frames
+        mel_dir = output_dir / "data_feats"
+        mel_dir.mkdir(parents=True, exist_ok=True)
+        mel_path = mel_dir / (utt_id + "_feats.npy")
+        np.save(mel_path, logmel)
+
+        if wav.size < num_frames * config.n_shift:
+            wav = np.pad(
+                wav, (0, num_frames * config.n_shift - wav.size),
+                mode="reflect")
+        else:
+            wav = wav[:num_frames * config.n_shift]
+        wave_dir = output_dir / "data_wave"
+        wave_dir.mkdir(parents=True, exist_ok=True)
+        wav_path = wave_dir / (utt_id + "_wave.npy")
+        # (num_samples, )
+        np.save(wav_path, wav)
+        # extract pitch and energy
+        if token_average == True:
+            f0 = pitch_extractor.get_pitch(
+                wav,
+                duration=np.array(durations),
+                use_token_averaged_f0=token_average)
+            if (f0 == 0).all():
+                return None
+            assert f0.shape[0] == len(durations)
+        else:
+            f0 = pitch_extractor.get_pitch(
+                wav, use_token_averaged_f0=token_average)
+            if (f0 == 0).all():
+                return None
+            f0 = f0[:num_frames]
+            assert f0.shape[0] == num_frames
+        f0_dir = output_dir / "data_pitch"
+        f0_dir.mkdir(parents=True, exist_ok=True)
+        f0_path = f0_dir / (utt_id + "_pitch.npy")
+        np.save(f0_path, f0)
+        if token_average == True:
+            energy = energy_extractor.get_energy(
+                wav,
+                duration=np.array(durations),
+                use_token_averaged_energy=token_average)
+            assert energy.shape[0] == len(durations)
+        else:
+            energy = energy_extractor.get_energy(
+                wav, use_token_averaged_energy=token_average)
+            energy = energy[:num_frames]
+            assert energy.shape[0] == num_frames
+
+        energy_dir = output_dir / "data_energy"
+        energy_dir.mkdir(parents=True, exist_ok=True)
+        energy_path = energy_dir / (utt_id + "_energy.npy")
+        np.save(energy_path, energy)
+        record = {
+            "utt_id": utt_id,
+            "phones": phones,
+            "text_lengths": len(phones),
+            "feats_lengths": num_frames,
+            "durations": durations,
+            "feats": str(mel_path),
+            "pitch": str(f0_path),
+            "energy": str(energy_path),
+            "wave": str(wav_path),
+            "speaker": speaker
+        }
+        if spk_emb_dir:
+            if speaker in os.listdir(spk_emb_dir):
+                embed_name = utt_id + ".npy"
+                embed_path = spk_emb_dir / speaker / embed_name
+                if embed_path.is_file():
+                    record["spk_emb"] = str(embed_path)
+                else:
+                    return None
+    return record
+
+
+def process_sentences(config,
+                      fps: List[Path],
+                      sentences: Dict,
+                      output_dir: Path,
+                      mel_extractor=None,
+                      pitch_extractor=None,
+                      energy_extractor=None,
+                      nprocs: int=1,
+                      cut_sil: bool=True,
+                      spk_emb_dir: Path=None,
+                      write_metadata_method: str='w',
+                      token_average: bool=True):
+    if nprocs == 1:
+        results = []
+        for fp in tqdm.tqdm(fps, total=len(fps)):
+            record = process_sentence(
+                config=config,
+                fp=fp,
+                sentences=sentences,
+                output_dir=output_dir,
+                mel_extractor=mel_extractor,
+                pitch_extractor=pitch_extractor,
+                energy_extractor=energy_extractor,
+                cut_sil=cut_sil,
+                spk_emb_dir=spk_emb_dir,
+                token_average=token_average)
+            if record:
+                results.append(record)
+    else:
+        with ThreadPoolExecutor(nprocs) as pool:
+            futures = []
+            with tqdm.tqdm(total=len(fps)) as progress:
+                for fp in fps:
+                    future = pool.submit(process_sentence, config, fp,
+                                         sentences, output_dir, mel_extractor,
+                                         pitch_extractor, energy_extractor,
+                                         cut_sil, spk_emb_dir)
+                    future.add_done_callback(lambda p: progress.update())
+                    futures.append(future)
+
+                results = []
+                for ft in futures:
+                    record = ft.result()
+                    if record:
+                        results.append(record)
+
+    results.sort(key=itemgetter("utt_id"))
+    with jsonlines.open(output_dir / "metadata.jsonl",
+                        write_metadata_method) as writer:
+        for item in results:
+            writer.write(item)
+    print("Done")
+
+
+def main():
+    # parse config and args
+    parser = argparse.ArgumentParser(
+        description="Preprocess audio and then extract features.")
+
+    parser.add_argument(
+        "--dataset",
+        default="baker",
+        type=str,
+        help="name of dataset, should in {baker, aishell3, ljspeech, vctk} now")
+
+    parser.add_argument(
+        "--rootdir", default=None, type=str, help="directory to dataset.")
+
+    parser.add_argument(
+        "--dumpdir",
+        type=str,
+        required=True,
+        help="directory to dump feature files.")
+    parser.add_argument(
+        "--dur-file", default=None, type=str, help="path to durations.txt.")
+
+    parser.add_argument("--config", type=str, help="fastspeech2 config file.")
+
+    parser.add_argument(
+        "--num-cpu", type=int, default=1, help="number of process.")
+
+    parser.add_argument(
+        "--cut-sil",
+        type=str2bool,
+        default=True,
+        help="whether cut sil in the edge of audio")
+
+    parser.add_argument(
+        "--spk_emb_dir",
+        default=None,
+        type=str,
+        help="directory to speaker embedding files.")
+
+    parser.add_argument(
+        "--write_metadata_method",
+        default="w",
+        type=str,
+        choices=["w", "a"],
+        help="How the metadata.jsonl file is written.")
+
+    parser.add_argument(
+        "--token_average",
+        type=str2bool,
+        default=False,
+        help="Average the energy and pitch accroding to durations")
+    args = parser.parse_args()
+
+    rootdir = Path(args.rootdir).expanduser()
+    dumpdir = Path(args.dumpdir).expanduser()
+    # use absolute path
+    dumpdir = dumpdir.resolve()
+    dumpdir.mkdir(parents=True, exist_ok=True)
+    dur_file = Path(args.dur_file).expanduser()
+
+    if args.spk_emb_dir:
+        spk_emb_dir = Path(args.spk_emb_dir).expanduser().resolve()
+    else:
+        spk_emb_dir = None
+
+    assert rootdir.is_dir()
+    assert dur_file.is_file()
+
+    with open(args.config, 'rt') as f:
+        config = CfgNode(yaml.safe_load(f))
+
+    sentences, speaker_set = get_phn_dur(dur_file)
+
+    merge_silence(sentences)
+    phone_id_map_path = dumpdir / "phone_id_map.txt"
+    speaker_id_map_path = dumpdir / "speaker_id_map.txt"
+    get_input_token(sentences, phone_id_map_path, args.dataset)
+    get_spk_id_map(speaker_set, speaker_id_map_path)
+
+    if args.dataset == "baker":
+        wav_files = sorted(list((rootdir / "Wave").rglob("*.wav")))
+        # split data into 3 sections
+        num_train = 9800
+        num_dev = 100
+        train_wav_files = wav_files[:num_train]
+        dev_wav_files = wav_files[num_train:num_train + num_dev]
+        test_wav_files = wav_files[num_train + num_dev:]
+    elif args.dataset == "aishell3":
+        sub_num_dev = 5
+        wav_dir = rootdir / "train" / "wav"
+        train_wav_files = []
+        dev_wav_files = []
+        test_wav_files = []
+        for speaker in os.listdir(wav_dir):
+            wav_files = sorted(list((wav_dir / speaker).rglob("*.wav")))
+            if len(wav_files) > 100:
+                train_wav_files += wav_files[:-sub_num_dev * 2]
+                dev_wav_files += wav_files[-sub_num_dev * 2:-sub_num_dev]
+                test_wav_files += wav_files[-sub_num_dev:]
+            else:
+                train_wav_files += wav_files
+    elif args.dataset == "canton":
+        sub_num_dev = 5
+        wav_dir = rootdir / "WAV"
+        train_wav_files = []
+        dev_wav_files = []
+        test_wav_files = []
+        for speaker in os.listdir(wav_dir):
+            wav_files = sorted(list((wav_dir / speaker).rglob("*.wav")))
+            if len(wav_files) > 100:
+                train_wav_files += wav_files[:-sub_num_dev * 2]
+                dev_wav_files += wav_files[-sub_num_dev * 2:-sub_num_dev]
+                test_wav_files += wav_files[-sub_num_dev:]
+            else:
+                train_wav_files += wav_files
+    elif args.dataset == "ljspeech":
+        wav_files = sorted(list((rootdir / "wavs").rglob("*.wav")))
+        # split data into 3 sections
+        num_train = 12900
+        num_dev = 100
+        train_wav_files = wav_files[:num_train]
+        dev_wav_files = wav_files[num_train:num_train + num_dev]
+        test_wav_files = wav_files[num_train + num_dev:]
+    elif args.dataset == "vctk":
+        sub_num_dev = 5
+        wav_dir = rootdir / "wav48_silence_trimmed"
+        train_wav_files = []
+        dev_wav_files = []
+        test_wav_files = []
+        for speaker in os.listdir(wav_dir):
+            wav_files = sorted(list((wav_dir / speaker).rglob("*_mic2.flac")))
+            if len(wav_files) > 100:
+                train_wav_files += wav_files[:-sub_num_dev * 2]
+                dev_wav_files += wav_files[-sub_num_dev * 2:-sub_num_dev]
+                test_wav_files += wav_files[-sub_num_dev:]
+            else:
+                train_wav_files += wav_files
+
+    else:
+        print("dataset should in {baker, aishell3, ljspeech, vctk} now!")
+
+    train_dump_dir = dumpdir / "train" / "raw"
+    train_dump_dir.mkdir(parents=True, exist_ok=True)
+    dev_dump_dir = dumpdir / "dev" / "raw"
+    dev_dump_dir.mkdir(parents=True, exist_ok=True)
+    test_dump_dir = dumpdir / "test" / "raw"
+    test_dump_dir.mkdir(parents=True, exist_ok=True)
+
+    # Extractor
+    mel_extractor = LogMelFBank(
+        sr=config.fs,
+        n_fft=config.n_fft,
+        hop_length=config.n_shift,
+        win_length=config.win_length,
+        window=config.window,
+        n_mels=config.n_mels,
+        fmin=config.fmin,
+        fmax=config.fmax)
+    pitch_extractor = Pitch(
+        sr=config.fs,
+        hop_length=config.n_shift,
+        f0min=config.f0min,
+        f0max=config.f0max)
+    energy_extractor = Energy(
+        n_fft=config.n_fft,
+        hop_length=config.n_shift,
+        win_length=config.win_length,
+        window=config.window)
+
+    # process for the 3 sections
+    if train_wav_files:
+        process_sentences(
+            config=config,
+            fps=train_wav_files,
+            sentences=sentences,
+            output_dir=train_dump_dir,
+            mel_extractor=mel_extractor,
+            pitch_extractor=pitch_extractor,
+            energy_extractor=energy_extractor,
+            nprocs=args.num_cpu,
+            cut_sil=args.cut_sil,
+            spk_emb_dir=spk_emb_dir,
+            write_metadata_method=args.write_metadata_method,
+            token_average=args.token_average)
+    if dev_wav_files:
+        process_sentences(
+            config=config,
+            fps=dev_wav_files,
+            sentences=sentences,
+            output_dir=dev_dump_dir,
+            mel_extractor=mel_extractor,
+            pitch_extractor=pitch_extractor,
+            energy_extractor=energy_extractor,
+            nprocs=args.num_cpu,
+            cut_sil=args.cut_sil,
+            spk_emb_dir=spk_emb_dir,
+            write_metadata_method=args.write_metadata_method,
+            token_average=args.token_average)
+    if test_wav_files:
+        process_sentences(
+            config=config,
+            fps=test_wav_files,
+            sentences=sentences,
+            output_dir=test_dump_dir,
+            mel_extractor=mel_extractor,
+            pitch_extractor=pitch_extractor,
+            energy_extractor=energy_extractor,
+            nprocs=args.num_cpu,
+            cut_sil=args.cut_sil,
+            spk_emb_dir=spk_emb_dir,
+            write_metadata_method=args.write_metadata_method,
+            token_average=args.token_average)
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/jets/synthesize.py

@@ -0,0 +1,153 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+from pathlib import Path
+
+import jsonlines
+import numpy as np
+import paddle
+import soundfile as sf
+import yaml
+from timer import timer
+from yacs.config import CfgNode
+
+from paddlespeech.t2s.datasets.data_table import DataTable
+from paddlespeech.t2s.models.jets import JETS
+from paddlespeech.t2s.utils import str2bool
+
+
+def evaluate(args):
+
+    # construct dataset for evaluation
+    with jsonlines.open(args.test_metadata, 'r') as reader:
+        test_metadata = list(reader)
+    # Init body.
+    with open(args.config) as f:
+        config = CfgNode(yaml.safe_load(f))
+
+    print("========Args========")
+    print(yaml.safe_dump(vars(args)))
+    print("========Config========")
+    print(config)
+
+    fields = ["utt_id", "text"]
+    converters = {}
+
+    spk_num = None
+    if args.speaker_dict is not None:
+        print("multiple speaker jets!")
+        with open(args.speaker_dict, 'rt') as f:
+            spk_id = [line.strip().split() for line in f.readlines()]
+        spk_num = len(spk_id)
+        fields += ["spk_id"]
+    elif args.voice_cloning:
+        print("Evaluating voice cloning!")
+        fields += ["spk_emb"]
+    else:
+        print("single speaker jets!")
+    print("spk_num:", spk_num)
+
+    test_dataset = DataTable(
+        data=test_metadata,
+        fields=fields,
+        converters=converters, )
+
+    with open(args.phones_dict, "r") as f:
+        phn_id = [line.strip().split() for line in f.readlines()]
+    vocab_size = len(phn_id)
+    print("vocab_size:", vocab_size)
+
+    odim = config.n_fft // 2 + 1
+    config["model"]["generator_params"]["spks"] = spk_num
+
+    jets = JETS(idim=vocab_size, odim=odim, **config["model"])
+    jets.set_state_dict(paddle.load(args.ckpt)["main_params"])
+    jets.eval()
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    N = 0
+    T = 0
+
+    for datum in test_dataset:
+        utt_id = datum["utt_id"]
+        phone_ids = paddle.to_tensor(datum["text"])
+        with timer() as t:
+            with paddle.no_grad():
+                spk_emb = None
+                spk_id = None
+                # multi speaker
+                if args.voice_cloning and "spk_emb" in datum:
+                    spk_emb = paddle.to_tensor(np.load(datum["spk_emb"]))
+                elif "spk_id" in datum:
+                    spk_id = paddle.to_tensor(datum["spk_id"])
+                out = jets.inference(
+                    text=phone_ids, sids=spk_id, spembs=spk_emb)
+            wav = out["wav"]
+            wav = wav.numpy()
+            N += wav.size
+            T += t.elapse
+            speed = wav.size / t.elapse
+            rtf = config.fs / speed
+        print(
+            f"{utt_id}, wave: {wav.size}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+        sf.write(str(output_dir / (utt_id + ".wav")), wav, samplerate=config.fs)
+        print(f"{utt_id} done!")
+    print(f"generation speed: {N / T}Hz, RTF: {config.fs / (N / T) }")
+
+
+def parse_args():
+    # parse args and config
+    parser = argparse.ArgumentParser(description="Synthesize with JETS")
+    # model
+    parser.add_argument(
+        '--config', type=str, default=None, help='Config of JETS.')
+    parser.add_argument(
+        '--ckpt', type=str, default=None, help='Checkpoint file of JETS.')
+    parser.add_argument(
+        "--phones_dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--speaker_dict", type=str, default=None, help="speaker id map file.")
+    parser.add_argument(
+        "--voice-cloning",
+        type=str2bool,
+        default=False,
+        help="whether training voice cloning model.")
+    # other
+    parser.add_argument(
+        "--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
+    parser.add_argument("--test_metadata", type=str, help="test metadata.")
+    parser.add_argument("--output_dir", type=str, help="output dir.")
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+
+    if args.ngpu == 0:
+        paddle.set_device("cpu")
+    elif args.ngpu > 0:
+        paddle.set_device("gpu")
+    else:
+        print("ngpu should >= 0 !")
+
+    evaluate(args)
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/jets/synthesize_e2e.py

@@ -0,0 +1,189 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+from pathlib import Path
+
+import paddle
+import soundfile as sf
+import yaml
+from timer import timer
+from yacs.config import CfgNode
+
+from paddlespeech.t2s.exps.syn_utils import am_to_static
+from paddlespeech.t2s.exps.syn_utils import get_frontend
+from paddlespeech.t2s.exps.syn_utils import get_sentences
+from paddlespeech.t2s.models.jets import JETS
+from paddlespeech.t2s.models.jets import JETSInference
+from paddlespeech.t2s.utils import str2bool
+
+
+def evaluate(args):
+    # Init body.
+    with open(args.config) as f:
+        config = CfgNode(yaml.safe_load(f))
+
+    print("========Args========")
+    print(yaml.safe_dump(vars(args)))
+    print("========Config========")
+    print(config)
+
+    sentences = get_sentences(text_file=args.text, lang=args.lang)
+
+    # frontend
+    frontend = get_frontend(lang=args.lang, phones_dict=args.phones_dict)
+    # acoustic model
+    am_name = args.am[:args.am.rindex('_')]
+    am_dataset = args.am[args.am.rindex('_') + 1:]
+
+    spk_num = None
+    if args.speaker_dict is not None:
+        print("multiple speaker jets!")
+        with open(args.speaker_dict, 'rt') as f:
+            spk_id = [line.strip().split() for line in f.readlines()]
+        spk_num = len(spk_id)
+    else:
+        print("single speaker jets!")
+    print("spk_num:", spk_num)
+
+    with open(args.phones_dict, "r") as f:
+        phn_id = [line.strip().split() for line in f.readlines()]
+    vocab_size = len(phn_id)
+    print("vocab_size:", vocab_size)
+
+    odim = config.n_fft // 2 + 1
+    config["model"]["generator_params"]["spks"] = spk_num
+
+    jets = JETS(idim=vocab_size, odim=odim, **config["model"])
+    jets.set_state_dict(paddle.load(args.ckpt)["main_params"])
+    jets.eval()
+
+    jets_inference = JETSInference(jets)
+    # whether dygraph to static
+    if args.inference_dir:
+        jets_inference = am_to_static(
+            am_inference=jets_inference,
+            am=args.am,
+            inference_dir=args.inference_dir,
+            speaker_dict=args.speaker_dict)
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    merge_sentences = False
+
+    N = 0
+    T = 0
+    for utt_id, sentence in sentences:
+        with timer() as t:
+            if args.lang == 'zh':
+                input_ids = frontend.get_input_ids(
+                    sentence, merge_sentences=merge_sentences)
+                phone_ids = input_ids["phone_ids"]
+            elif args.lang == 'en':
+                input_ids = frontend.get_input_ids(
+                    sentence, merge_sentences=merge_sentences)
+                phone_ids = input_ids["phone_ids"]
+            else:
+                print("lang should in {'zh', 'en'}!")
+            with paddle.no_grad():
+                flags = 0
+                for i in range(len(phone_ids)):
+                    part_phone_ids = phone_ids[i]
+                    spk_id = None
+                    if am_dataset in {"aishell3",
+                                      "vctk"} and spk_num is not None:
+                        spk_id = paddle.to_tensor(args.spk_id)
+                        wav = jets_inference(part_phone_ids, spk_id)
+                    else:
+                        wav = jets_inference(part_phone_ids)
+                    if flags == 0:
+                        wav_all = wav
+                        flags = 1
+                    else:
+                        wav_all = paddle.concat([wav_all, wav])
+        wav = wav_all.numpy()
+        N += wav.size
+        T += t.elapse
+        speed = wav.size / t.elapse
+        rtf = config.fs / speed
+        print(
+            f"{utt_id}, wave: {wav.shape}, time: {t.elapse}s, Hz: {speed}, RTF: {rtf}."
+        )
+        sf.write(str(output_dir / (utt_id + ".wav")), wav, samplerate=config.fs)
+        print(f"{utt_id} done!")
+    print(f"generation speed: {N / T}Hz, RTF: {config.fs / (N / T) }")
+
+
+def parse_args():
+    # parse args and config 
+    parser = argparse.ArgumentParser(description="Synthesize with JETS")
+
+    # model
+    parser.add_argument(
+        '--config', type=str, default=None, help='Config of JETS.')
+    parser.add_argument(
+        '--ckpt', type=str, default=None, help='Checkpoint file of JETS.')
+    parser.add_argument(
+        "--phones_dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--speaker_dict", type=str, default=None, help="speaker id map file.")
+    parser.add_argument(
+        '--spk_id',
+        type=int,
+        default=0,
+        help='spk id for multi speaker acoustic model')
+    # other
+    parser.add_argument(
+        '--lang',
+        type=str,
+        default='zh',
+        help='Choose model language. zh or en')
+
+    parser.add_argument(
+        "--inference_dir",
+        type=str,
+        default=None,
+        help="dir to save inference models")
+    parser.add_argument(
+        "--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
+    parser.add_argument(
+        "--text",
+        type=str,
+        help="text to synthesize, a 'utt_id sentence' pair per line.")
+    parser.add_argument("--output_dir", type=str, help="output dir.")
+
+    parser.add_argument(
+        '--am',
+        type=str,
+        default='jets_csmsc',
+        help='Choose acoustic model type of tts task.')
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = parse_args()
+
+    if args.ngpu == 0:
+        paddle.set_device("cpu")
+    elif args.ngpu > 0:
+        paddle.set_device("gpu")
+    else:
+        print("ngpu should >= 0 !")
+
+    evaluate(args)
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/jets/train.py

@@ -0,0 +1,305 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import argparse
+import logging
+import os
+import shutil
+from pathlib import Path
+
+import jsonlines
+import numpy as np
+import paddle
+import yaml
+from paddle import DataParallel
+from paddle import distributed as dist
+from paddle.io import DataLoader
+from paddle.optimizer import AdamW
+from yacs.config import CfgNode
+
+from paddlespeech.t2s.datasets.am_batch_fn import jets_multi_spk_batch_fn
+from paddlespeech.t2s.datasets.am_batch_fn import jets_single_spk_batch_fn
+from paddlespeech.t2s.datasets.data_table import DataTable
+from paddlespeech.t2s.datasets.sampler import ErnieSATSampler
+from paddlespeech.t2s.models.jets import JETS
+from paddlespeech.t2s.models.jets import JETSEvaluator
+from paddlespeech.t2s.models.jets import JETSUpdater
+from paddlespeech.t2s.modules.losses import DiscriminatorAdversarialLoss
+from paddlespeech.t2s.modules.losses import FeatureMatchLoss
+from paddlespeech.t2s.modules.losses import ForwardSumLoss
+from paddlespeech.t2s.modules.losses import GeneratorAdversarialLoss
+from paddlespeech.t2s.modules.losses import MelSpectrogramLoss
+from paddlespeech.t2s.modules.losses import VarianceLoss
+from paddlespeech.t2s.training.extensions.snapshot import Snapshot
+from paddlespeech.t2s.training.extensions.visualizer import VisualDL
+from paddlespeech.t2s.training.optimizer import scheduler_classes
+from paddlespeech.t2s.training.seeding import seed_everything
+from paddlespeech.t2s.training.trainer import Trainer
+from paddlespeech.t2s.utils import str2bool
+
+
+def train_sp(args, config):
+    # decides device type and whether to run in parallel
+    # setup running environment correctly
+    world_size = paddle.distributed.get_world_size()
+    if (not paddle.is_compiled_with_cuda()) or args.ngpu == 0:
+        paddle.set_device("cpu")
+    else:
+        paddle.set_device("gpu")
+        if world_size > 1:
+            paddle.distributed.init_parallel_env()
+
+    # set the random seed, it is a must for multiprocess training
+    seed_everything(config.seed)
+
+    print(
+        f"rank: {dist.get_rank()}, pid: {os.getpid()}, parent_pid: {os.getppid()}",
+    )
+
+    # dataloader has been too verbose
+    logging.getLogger("DataLoader").disabled = True
+
+    fields = [
+        "text", "text_lengths", "feats", "feats_lengths", "wave", "durations",
+        "pitch", "energy"
+    ]
+
+    converters = {
+        "wave": np.load,
+        "feats": np.load,
+        "pitch": np.load,
+        "energy": np.load,
+    }
+    spk_num = None
+    if args.speaker_dict is not None:
+        print("multiple speaker jets!")
+        collate_fn = jets_multi_spk_batch_fn
+        with open(args.speaker_dict, 'rt', encoding='utf-8') as f:
+            spk_id = [line.strip().split() for line in f.readlines()]
+        spk_num = len(spk_id)
+        fields += ["spk_id"]
+    elif args.voice_cloning:
+        print("Training voice cloning!")
+        collate_fn = jets_multi_spk_batch_fn
+        fields += ["spk_emb"]
+        converters["spk_emb"] = np.load
+    else:
+        print("single speaker jets!")
+        collate_fn = jets_single_spk_batch_fn
+    print("spk_num:", spk_num)
+
+    # construct dataset for training and validation
+    with jsonlines.open(args.train_metadata, 'r') as reader:
+        train_metadata = list(reader)
+    train_dataset = DataTable(
+        data=train_metadata,
+        fields=fields,
+        converters=converters, )
+    with jsonlines.open(args.dev_metadata, 'r') as reader:
+        dev_metadata = list(reader)
+    dev_dataset = DataTable(
+        data=dev_metadata,
+        fields=fields,
+        converters=converters, )
+
+    # collate function and dataloader
+    train_sampler = ErnieSATSampler(
+        train_dataset,
+        batch_size=config.batch_size,
+        shuffle=False,
+        drop_last=True)
+    dev_sampler = ErnieSATSampler(
+        dev_dataset,
+        batch_size=config.batch_size,
+        shuffle=False,
+        drop_last=False)
+    print("samplers done!")
+
+    train_dataloader = DataLoader(
+        train_dataset,
+        batch_sampler=train_sampler,
+        collate_fn=collate_fn,
+        num_workers=config.num_workers)
+
+    dev_dataloader = DataLoader(
+        dev_dataset,
+        batch_sampler=dev_sampler,
+        collate_fn=collate_fn,
+        num_workers=config.num_workers)
+    print("dataloaders done!")
+
+    with open(args.phones_dict, 'rt', encoding='utf-8') as f:
+        phn_id = [line.strip().split() for line in f.readlines()]
+    vocab_size = len(phn_id)
+    print("vocab_size:", vocab_size)
+
+    odim = config.n_mels
+    config["model"]["generator_params"]["spks"] = spk_num
+    model = JETS(idim=vocab_size, odim=odim, **config["model"])
+    gen_parameters = model.generator.parameters()
+    dis_parameters = model.discriminator.parameters()
+    if world_size > 1:
+        model = DataParallel(model)
+        gen_parameters = model._layers.generator.parameters()
+        dis_parameters = model._layers.discriminator.parameters()
+
+    print("model done!")
+
+    # loss
+    criterion_mel = MelSpectrogramLoss(
+        **config["mel_loss_params"], )
+    criterion_feat_match = FeatureMatchLoss(
+        **config["feat_match_loss_params"], )
+    criterion_gen_adv = GeneratorAdversarialLoss(
+        **config["generator_adv_loss_params"], )
+    criterion_dis_adv = DiscriminatorAdversarialLoss(
+        **config["discriminator_adv_loss_params"], )
+    criterion_var = VarianceLoss()
+    criterion_forwardsum = ForwardSumLoss()
+
+    print("criterions done!")
+
+    lr_schedule_g = scheduler_classes[config["generator_scheduler"]](
+        **config["generator_scheduler_params"])
+    optimizer_g = AdamW(
+        learning_rate=lr_schedule_g,
+        parameters=gen_parameters,
+        **config["generator_optimizer_params"])
+
+    lr_schedule_d = scheduler_classes[config["discriminator_scheduler"]](
+        **config["discriminator_scheduler_params"])
+    optimizer_d = AdamW(
+        learning_rate=lr_schedule_d,
+        parameters=dis_parameters,
+        **config["discriminator_optimizer_params"])
+
+    print("optimizers done!")
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    if dist.get_rank() == 0:
+        config_name = args.config.split("/")[-1]
+        # copy conf to output_dir
+        shutil.copyfile(args.config, output_dir / config_name)
+
+    updater = JETSUpdater(
+        model=model,
+        optimizers={
+            "generator": optimizer_g,
+            "discriminator": optimizer_d,
+        },
+        criterions={
+            "mel": criterion_mel,
+            "feat_match": criterion_feat_match,
+            "gen_adv": criterion_gen_adv,
+            "dis_adv": criterion_dis_adv,
+            "var": criterion_var,
+            "forwardsum": criterion_forwardsum,
+        },
+        schedulers={
+            "generator": lr_schedule_g,
+            "discriminator": lr_schedule_d,
+        },
+        dataloader=train_dataloader,
+        lambda_adv=config.lambda_adv,
+        lambda_mel=config.lambda_mel,
+        lambda_feat_match=config.lambda_feat_match,
+        lambda_var=config.lambda_var,
+        lambda_align=config.lambda_align,
+        generator_first=config.generator_first,
+        use_alignment_module=config.use_alignment_module,
+        output_dir=output_dir)
+
+    evaluator = JETSEvaluator(
+        model=model,
+        criterions={
+            "mel": criterion_mel,
+            "feat_match": criterion_feat_match,
+            "gen_adv": criterion_gen_adv,
+            "dis_adv": criterion_dis_adv,
+            "var": criterion_var,
+            "forwardsum": criterion_forwardsum,
+        },
+        dataloader=dev_dataloader,
+        lambda_adv=config.lambda_adv,
+        lambda_mel=config.lambda_mel,
+        lambda_feat_match=config.lambda_feat_match,
+        lambda_var=config.lambda_var,
+        lambda_align=config.lambda_align,
+        generator_first=config.generator_first,
+        use_alignment_module=config.use_alignment_module,
+        output_dir=output_dir)
+
+    trainer = Trainer(
+        updater,
+        stop_trigger=(config.train_max_steps, "iteration"),
+        out=output_dir)
+
+    if dist.get_rank() == 0:
+        trainer.extend(
+            evaluator, trigger=(config.eval_interval_steps, 'iteration'))
+        trainer.extend(VisualDL(output_dir), trigger=(1, 'iteration'))
+    trainer.extend(
+        Snapshot(max_size=config.num_snapshots),
+        trigger=(config.save_interval_steps, 'iteration'))
+
+    print("Trainer Done!")
+    trainer.run()
+
+
+def main():
+    # parse args and config and redirect to train_sp
+
+    parser = argparse.ArgumentParser(description="Train a JETS model.")
+    parser.add_argument("--config", type=str, help="JETS config file")
+    parser.add_argument("--train-metadata", type=str, help="training data.")
+    parser.add_argument("--dev-metadata", type=str, help="dev data.")
+    parser.add_argument("--output-dir", type=str, help="output dir.")
+    parser.add_argument(
+        "--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
+    parser.add_argument(
+        "--phones-dict", type=str, default=None, help="phone vocabulary file.")
+    parser.add_argument(
+        "--speaker-dict",
+        type=str,
+        default=None,
+        help="speaker id map file for multiple speaker model.")
+
+    parser.add_argument(
+        "--voice-cloning",
+        type=str2bool,
+        default=False,
+        help="whether training voice cloning model.")
+
+    args = parser.parse_args()
+
+    with open(args.config, 'rt') as f:
+        config = CfgNode(yaml.safe_load(f))
+
+    print("========Args========")
+    print(yaml.safe_dump(vars(args)))
+    print("========Config========")
+    print(config)
+    print(
+        f"master see the word size: {dist.get_world_size()}, from pid: {os.getpid()}"
+    )
+
+    # dispatch
+    if args.ngpu > 1:
+        dist.spawn(train_sp, (args, config), nprocs=args.ngpu)
+    else:
+        train_sp(args, config)
+
+
+if __name__ == "__main__":
+    main()

paddlespeech/t2s/exps/syn_utils.py

@@ -506,7 +506,7 @@ def am_to_static(am_inference,
         am_inference = jit.to_static(
             am_inference, input_spec=[InputSpec([-1], dtype=paddle.int64)])
 
-    elif am_name == 'vits':
+    elif am_name == 'vits' or am_name == 'jets':
         if am_dataset in {"aishell3", "vctk"} and speaker_dict is not None:
             am_inference = jit.to_static(
                 am_inference,

paddlespeech/t2s/models/jets/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from .jets import *
+from .jets_updater import *

paddlespeech/t2s/models/jets/alignments.py

@@ -0,0 +1,182 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Generator module in JETS.
+
+This code is based on https://github.com/imdanboy/jets.
+
+"""
+import numpy as np
+import paddle
+import paddle.nn.functional as F
+from numba import jit
+from paddle import nn
+
+from paddlespeech.t2s.modules.masked_fill import masked_fill
+
+
+class AlignmentModule(nn.Layer):
+    """Alignment Learning Framework proposed for parallel TTS models in:
+    https://arxiv.org/abs/2108.10447
+    """
+
+    def __init__(self, adim, odim):
+        super().__init__()
+        self.t_conv1 = nn.Conv1D(adim, adim, kernel_size=3, padding=1)
+        self.t_conv2 = nn.Conv1D(adim, adim, kernel_size=1, padding=0)
+
+        self.f_conv1 = nn.Conv1D(odim, adim, kernel_size=3, padding=1)
+        self.f_conv2 = nn.Conv1D(adim, adim, kernel_size=3, padding=1)
+        self.f_conv3 = nn.Conv1D(adim, adim, kernel_size=1, padding=0)
+
+    def forward(self, text, feats, x_masks=None):
+        """
+        Args:
+            text (Tensor): Batched text embedding (B, T_text, adim)
+            feats (Tensor): Batched acoustic feature (B, T_feats, odim)
+            x_masks (Tensor): Mask tensor (B, T_text)
+
+        Returns:
+            Tensor: log probability of attention matrix (B, T_feats, T_text)
+        """
+
+        text = text.transpose((0, 2, 1))
+        text = F.relu(self.t_conv1(text))
+        text = self.t_conv2(text)
+        text = text.transpose((0, 2, 1))
+
+        feats = feats.transpose((0, 2, 1))
+        feats = F.relu(self.f_conv1(feats))
+        feats = F.relu(self.f_conv2(feats))
+        feats = self.f_conv3(feats)
+        feats = feats.transpose((0, 2, 1))
+
+        dist = feats.unsqueeze(2) - text.unsqueeze(1)
+        dist = paddle.linalg.norm(dist, p=2, axis=3)
+        score = -dist
+
+        if x_masks is not None:
+            x_masks = x_masks.unsqueeze(-2)
+            score = masked_fill(score, x_masks, -np.inf)
+        log_p_attn = F.log_softmax(score, axis=-1)
+        return log_p_attn, score
+
+
+@jit(nopython=True)
+def _monotonic_alignment_search(log_p_attn):
+    # https://arxiv.org/abs/2005.11129
+    T_mel = log_p_attn.shape[0]
+    T_inp = log_p_attn.shape[1]
+    Q = np.full((T_inp, T_mel), fill_value=-np.inf)
+
+    log_prob = log_p_attn.transpose(1, 0)  # -> (T_inp,T_mel)
+    # 1.  Q <- init first row for all j
+    for j in range(T_mel):
+        Q[0, j] = log_prob[0, :j + 1].sum()
+
+    # 2. 
+    for j in range(1, T_mel):
+        for i in range(1, min(j + 1, T_inp)):
+            Q[i, j] = max(Q[i - 1, j - 1], Q[i, j - 1]) + log_prob[i, j]
+
+    # 3.
+    A = np.full((T_mel, ), fill_value=T_inp - 1)
+    for j in range(T_mel - 2, -1, -1):  # T_mel-2, ..., 0
+        # 'i' in {A[j+1]-1, A[j+1]}
+        i_a = A[j + 1] - 1
+        i_b = A[j + 1]
+        if i_b == 0:
+            argmax_i = 0
+        elif Q[i_a, j] >= Q[i_b, j]:
+            argmax_i = i_a
+        else:
+            argmax_i = i_b
+        A[j] = argmax_i
+    return A
+
+
+def viterbi_decode(log_p_attn, text_lengths, feats_lengths):
+    """
+    Args:
+        log_p_attn (Tensor): 
+            Batched log probability of attention matrix (B, T_feats, T_text)
+        text_lengths (Tensor): 
+            Text length tensor (B,)
+        feats_legnths (Tensor): 
+            Feature length tensor (B,)
+    Returns:
+        Tensor: 
+            Batched token duration extracted from `log_p_attn` (B,T_text)
+        Tensor: 
+            binarization loss tensor ()
+    """
+    B = log_p_attn.shape[0]
+    T_text = log_p_attn.shape[2]
+    device = log_p_attn.place
+
+    bin_loss = 0
+    ds = paddle.zeros((B, T_text), dtype="int32")
+    for b in range(B):
+        cur_log_p_attn = log_p_attn[b, :feats_lengths[b], :text_lengths[b]]
+        viterbi = _monotonic_alignment_search(cur_log_p_attn.numpy())
+        _ds = np.bincount(viterbi)
+        ds[b, :len(_ds)] = paddle.to_tensor(
+            _ds, place=device, dtype="int32")  
+
+        t_idx = paddle.arange(feats_lengths[b])
+        bin_loss = bin_loss - cur_log_p_attn[t_idx, viterbi].mean()
+    bin_loss = bin_loss / B
+    return ds, bin_loss
+
+
+@jit(nopython=True)
+def _average_by_duration(ds, xs, text_lengths, feats_lengths):
+    B = ds.shape[0]
+    # xs_avg = np.zeros_like(ds)
+    xs_avg = np.zeros(shape=ds.shape, dtype=np.float32)
+    ds = ds.astype(np.int32)
+    for b in range(B):
+        t_text = text_lengths[b]
+        t_feats = feats_lengths[b]
+        d = ds[b, :t_text]
+        d_cumsum = d.cumsum()
+        d_cumsum = [0] + list(d_cumsum)
+        x = xs[b, :t_feats]
+        for n, (start, end) in enumerate(zip(d_cumsum[:-1], d_cumsum[1:])):
+            if len(x[start:end]) != 0:
+                xs_avg[b, n] = x[start:end].mean()
+            else:
+                xs_avg[b, n] = 0
+    return xs_avg
+
+
+def average_by_duration(ds, xs, text_lengths, feats_lengths):
+    """
+    Args:
+        ds (Tensor): 
+            Batched token duration (B,T_text)
+        xs (Tensor): 
+            Batched feature sequences to be averaged (B,T_feats)
+        text_lengths (Tensor): 
+            Text length tensor (B,)
+        feats_lengths (Tensor): 
+            Feature length tensor (B,)
+    Returns:
+        Tensor: Batched feature averaged according to the token duration (B, T_text)
+    """
+    device = ds.place
+    args = [ds, xs, text_lengths, feats_lengths]
+    args = [arg.numpy() for arg in args]
+    xs_avg = _average_by_duration(*args)
+    xs_avg = paddle.to_tensor(xs_avg, place=device)
+    return xs_avg

paddlespeech/t2s/models/jets/generator.py

@@ -0,0 +1,897 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Generator module in JETS.
+
+This code is based on https://github.com/imdanboy/jets.
+
+"""
+import logging
+import math
+from typing import Any
+from typing import Dict
+from typing import List
+from typing import Optional
+from typing import Sequence
+from typing import Tuple
+
+import numpy as np
+import paddle
+from paddle import nn
+from typeguard import check_argument_types
+
+from paddlespeech.t2s.models.hifigan import HiFiGANGenerator
+from paddlespeech.t2s.models.jets.alignments import AlignmentModule
+from paddlespeech.t2s.models.jets.alignments import average_by_duration
+from paddlespeech.t2s.models.jets.alignments import viterbi_decode
+from paddlespeech.t2s.models.jets.length_regulator import GaussianUpsampling
+from paddlespeech.t2s.modules.nets_utils import get_random_segments
+from paddlespeech.t2s.modules.nets_utils import initialize
+from paddlespeech.t2s.modules.nets_utils import make_non_pad_mask
+from paddlespeech.t2s.modules.nets_utils import make_pad_mask
+from paddlespeech.t2s.modules.predictor.duration_predictor import DurationPredictor
+from paddlespeech.t2s.modules.predictor.length_regulator import LengthRegulator
+from paddlespeech.t2s.modules.predictor.variance_predictor import VariancePredictor
+from paddlespeech.t2s.modules.style_encoder import StyleEncoder
+from paddlespeech.t2s.modules.transformer.embedding import PositionalEncoding
+from paddlespeech.t2s.modules.transformer.embedding import ScaledPositionalEncoding
+from paddlespeech.t2s.modules.transformer.encoder import ConformerEncoder
+from paddlespeech.t2s.modules.transformer.encoder import TransformerEncoder
+
+
+class JETSGenerator(nn.Layer):
+    """Generator module in JETS.
+    """
+
+    def __init__(
+            self,
+            idim: int,
+            odim: int,
+            adim: int=256,
+            aheads: int=2,
+            elayers: int=4,
+            eunits: int=1024,
+            dlayers: int=4,
+            dunits: int=1024,
+            positionwise_layer_type: str="conv1d",
+            positionwise_conv_kernel_size: int=1,
+            use_scaled_pos_enc: bool=True,
+            use_batch_norm: bool=True,
+            encoder_normalize_before: bool=True,
+            decoder_normalize_before: bool=True,
+            encoder_concat_after: bool=False,
+            decoder_concat_after: bool=False,
+            reduction_factor: int=1,
+            encoder_type: str="transformer",
+            decoder_type: str="transformer",
+            transformer_enc_dropout_rate: float=0.1,
+            transformer_enc_positional_dropout_rate: float=0.1,
+            transformer_enc_attn_dropout_rate: float=0.1,
+            transformer_dec_dropout_rate: float=0.1,
+            transformer_dec_positional_dropout_rate: float=0.1,
+            transformer_dec_attn_dropout_rate: float=0.1,
+            transformer_activation_type: str="relu",
+            # only for conformer
+            conformer_rel_pos_type: str="legacy",
+            conformer_pos_enc_layer_type: str="rel_pos",
+            conformer_self_attn_layer_type: str="rel_selfattn",
+            conformer_activation_type: str="swish",
+            use_macaron_style_in_conformer: bool=True,
+            use_cnn_in_conformer: bool=True,
+            zero_triu: bool=False,
+            conformer_enc_kernel_size: int=7,
+            conformer_dec_kernel_size: int=31,
+            # duration predictor
+            duration_predictor_layers: int=2,
+            duration_predictor_chans: int=384,
+            duration_predictor_kernel_size: int=3,
+            duration_predictor_dropout_rate: float=0.1,
+            # energy predictor
+            energy_predictor_layers: int=2,
+            energy_predictor_chans: int=384,
+            energy_predictor_kernel_size: int=3,
+            energy_predictor_dropout: float=0.5,
+            energy_embed_kernel_size: int=9,
+            energy_embed_dropout: float=0.5,
+            stop_gradient_from_energy_predictor: bool=False,
+            # pitch predictor
+            pitch_predictor_layers: int=2,
+            pitch_predictor_chans: int=384,
+            pitch_predictor_kernel_size: int=3,
+            pitch_predictor_dropout: float=0.5,
+            pitch_embed_kernel_size: int=9,
+            pitch_embed_dropout: float=0.5,
+            stop_gradient_from_pitch_predictor: bool=False,
+            # extra embedding related
+            spks: Optional[int]=None,
+            langs: Optional[int]=None,
+            spk_embed_dim: Optional[int]=None,
+            spk_embed_integration_type: str="add",
+            use_gst: bool=False,
+            gst_tokens: int=10,
+            gst_heads: int=4,
+            gst_conv_layers: int=6,
+            gst_conv_chans_list: Sequence[int]=(32, 32, 64, 64, 128, 128),
+            gst_conv_kernel_size: int=3,
+            gst_conv_stride: int=2,
+            gst_gru_layers: int=1,
+            gst_gru_units: int=128,
+            # training related
+            init_type: str="xavier_uniform",
+            init_enc_alpha: float=1.0,
+            init_dec_alpha: float=1.0,
+            use_masking: bool=False,
+            use_weighted_masking: bool=False,
+            segment_size: int=64,
+            # hifigan generator
+            generator_out_channels: int=1,
+            generator_channels: int=512,
+            generator_global_channels: int=-1,
+            generator_kernel_size: int=7,
+            generator_upsample_scales: List[int]=[8, 8, 2, 2],
+            generator_upsample_kernel_sizes: List[int]=[16, 16, 4, 4],
+            generator_resblock_kernel_sizes: List[int]=[3, 7, 11],
+            generator_resblock_dilations: List[List[int]]=[[1, 3, 5], [1, 3, 5],
+                                                           [1, 3, 5]],
+            generator_use_additional_convs: bool=True,
+            generator_bias: bool=True,
+            generator_nonlinear_activation: str="LeakyReLU",
+            generator_nonlinear_activation_params: Dict[
+                str, Any]={"negative_slope": 0.1},
+            generator_use_weight_norm: bool=True, ):
+        """Initialize JETS generator module.
+
+        Args:
+            idim (int): 
+                Dimension of the inputs.
+            odim (int): 
+                Dimension of the outputs.
+            adim (int): 
+                Attention dimension.
+            aheads (int): 
+                Number of attention heads.
+            elayers (int): 
+                Number of encoder layers.
+            eunits (int): 
+                Number of encoder hidden units.
+            dlayers (int): 
+                Number of decoder layers.
+            dunits (int): 
+                Number of decoder hidden units.
+            use_scaled_pos_enc (bool): 
+                Whether to use trainable scaled pos encoding.
+            use_batch_norm (bool): 
+                Whether to use batch normalization in encoder prenet.
+            encoder_normalize_before (bool): 
+                Whether to apply layernorm layer before encoder block.
+            decoder_normalize_before (bool): 
+                Whether to apply layernorm layer before decoder block.
+            encoder_concat_after (bool): 
+                Whether to concatenate attention layer's input and output in encoder.
+            decoder_concat_after (bool): 
+                Whether to concatenate attention layer's input and output in decoder.
+            reduction_factor (int): 
+                Reduction factor.
+            encoder_type (str): 
+                Encoder type ("transformer" or "conformer").
+            decoder_type (str): 
+                Decoder type ("transformer" or "conformer").
+            transformer_enc_dropout_rate (float): 
+                Dropout rate in encoder except attention and positional encoding.
+            transformer_enc_positional_dropout_rate (float): 
+                Dropout rate after encoder positional encoding.
+            transformer_enc_attn_dropout_rate (float): 
+                Dropout rate in encoder self-attention module.
+            transformer_dec_dropout_rate (float): 
+                Dropout rate in decoder except attention & positional encoding.
+            transformer_dec_positional_dropout_rate (float): 
+                Dropout rate after decoder positional encoding.
+            transformer_dec_attn_dropout_rate (float): 
+                Dropout rate in decoder self-attention module.
+            conformer_rel_pos_type (str): 
+                Relative pos encoding type in conformer.
+            conformer_pos_enc_layer_type (str): 
+                Pos encoding layer type in conformer.
+            conformer_self_attn_layer_type (str): 
+                Self-attention layer type in conformer
+            conformer_activation_type (str): 
+                Activation function type in conformer.
+            use_macaron_style_in_conformer: 
+                Whether to use macaron style FFN.
+            use_cnn_in_conformer: 
+                Whether to use CNN in conformer.
+            zero_triu: 
+                Whether to use zero triu in relative self-attention module.
+            conformer_enc_kernel_size: 
+                Kernel size of encoder conformer.
+            conformer_dec_kernel_size: 
+                Kernel size of decoder conformer.
+            duration_predictor_layers (int): 
+                Number of duration predictor layers.
+            duration_predictor_chans (int): 
+                Number of duration predictor channels.
+            duration_predictor_kernel_size (int): 
+                Kernel size of duration predictor.
+            duration_predictor_dropout_rate (float): 
+                Dropout rate in duration predictor.
+            pitch_predictor_layers (int): 
+                Number of pitch predictor layers.
+            pitch_predictor_chans (int): 
+                Number of pitch predictor channels.
+            pitch_predictor_kernel_size (int): 
+                Kernel size of pitch predictor.
+            pitch_predictor_dropout_rate (float): 
+                Dropout rate in pitch predictor.
+            pitch_embed_kernel_size (float): 
+                Kernel size of pitch embedding.
+            pitch_embed_dropout_rate (float): 
+                Dropout rate for pitch embedding.
+            stop_gradient_from_pitch_predictor: 
+                Whether to stop gradient from pitch predictor to encoder.
+            energy_predictor_layers (int): 
+                Number of energy predictor layers.
+            energy_predictor_chans (int): 
+                Number of energy predictor channels.
+            energy_predictor_kernel_size (int): 
+                Kernel size of energy predictor.
+            energy_predictor_dropout_rate (float): 
+                Dropout rate in energy predictor.
+            energy_embed_kernel_size (float): 
+                Kernel size of energy embedding.
+            energy_embed_dropout_rate (float): 
+                Dropout rate for energy embedding.
+            stop_gradient_from_energy_predictor: 
+                Whether to stop gradient from energy predictor to encoder.
+            spks (Optional[int]): 
+                Number of speakers. If set to > 1, assume that the sids will be provided as the input and use sid embedding layer.
+            langs (Optional[int]): 
+                Number of languages. If set to > 1, assume that the lids will be provided as the input and use sid embedding layer.
+            spk_embed_dim (Optional[int]): 
+                Speaker embedding dimension. If set to > 0, assume that spembs will be provided as the input.
+            spk_embed_integration_type: 
+                How to integrate speaker embedding.
+            use_gst (str): 
+                Whether to use global style token.
+            gst_tokens (int): 
+                The number of GST embeddings.
+            gst_heads (int): 
+                The number of heads in GST multihead attention.
+            gst_conv_layers (int): 
+                The number of conv layers in GST.
+            gst_conv_chans_list: (Sequence[int]):
+                List of the number of channels of conv layers in GST.
+            gst_conv_kernel_size (int): 
+                Kernel size of conv layers in GST.
+            gst_conv_stride (int): 
+                Stride size of conv layers in GST.
+            gst_gru_layers (int): 
+                The number of GRU layers in GST.
+            gst_gru_units (int): 
+                The number of GRU units in GST.
+            init_type (str): 
+                How to initialize transformer parameters.
+            init_enc_alpha (float): 
+                Initial value of alpha in scaled pos encoding of the encoder.
+            init_dec_alpha (float): 
+                Initial value of alpha in scaled pos encoding of the decoder.
+            use_masking (bool): 
+                Whether to apply masking for padded part in loss calculation.
+            use_weighted_masking (bool): 
+                Whether to apply weighted masking in loss calculation.
+            segment_size (int): 
+                Segment size for random windowed discriminator
+            generator_out_channels (int): 
+                Number of output channels.
+            generator_channels (int): 
+                Number of hidden representation channels.
+            generator_global_channels (int): 
+                Number of global conditioning channels.
+            generator_kernel_size (int): 
+                Kernel size of initial and final conv layer.
+            generator_upsample_scales (List[int]): 
+                List of upsampling scales.
+            generator_upsample_kernel_sizes (List[int]): 
+                List of kernel sizes for upsample layers.
+            generator_resblock_kernel_sizes (List[int]): 
+                List of kernel sizes for residual blocks.
+            generator_resblock_dilations (List[List[int]]): 
+                List of list of dilations for residual blocks.
+            generator_use_additional_convs (bool): 
+                Whether to use additional conv layers in residual blocks.
+            generator_bias (bool): 
+                Whether to add bias parameter in convolution layers.
+            generator_nonlinear_activation (str): 
+                Activation function module name.
+            generator_nonlinear_activation_params (Dict[str, Any]): 
+                Hyperparameters for activation function.
+            generator_use_weight_norm (bool): 
+                Whether to use weight norm. If set to true, it will be applied to all of the conv layers.
+
+        """
+        super().__init__()
+        self.segment_size = segment_size
+        self.upsample_factor = int(np.prod(generator_upsample_scales))
+        self.idim = idim
+        self.odim = odim
+        self.reduction_factor = reduction_factor
+        self.encoder_type = encoder_type
+        self.decoder_type = decoder_type
+        self.stop_gradient_from_pitch_predictor = stop_gradient_from_pitch_predictor
+        self.stop_gradient_from_energy_predictor = stop_gradient_from_energy_predictor
+        self.use_scaled_pos_enc = use_scaled_pos_enc
+        self.use_gst = use_gst
+
+        # use idx 0 as padding idx
+        self.padding_idx = 0
+
+        # get positional encoding layer type
+        transformer_pos_enc_layer_type = "scaled_abs_pos" if self.use_scaled_pos_enc else "abs_pos"
+
+        # check relative positional encoding compatibility
+        if "conformer" in [encoder_type, decoder_type]:
+            if conformer_rel_pos_type == "legacy":
+                if conformer_pos_enc_layer_type == "rel_pos":
+                    conformer_pos_enc_layer_type = "legacy_rel_pos"
+                    logging.warning(
+                        "Fallback to conformer_pos_enc_layer_type = 'legacy_rel_pos' "
+                        "due to the compatibility. If you want to use the new one, "
+                        "please use conformer_pos_enc_layer_type = 'latest'.")
+                if conformer_self_attn_layer_type == "rel_selfattn":
+                    conformer_self_attn_layer_type = "legacy_rel_selfattn"
+                    logging.warning(
+                        "Fallback to "
+                        "conformer_self_attn_layer_type = 'legacy_rel_selfattn' "
+                        "due to the compatibility. If you want to use the new one, "
+                        "please use conformer_pos_enc_layer_type = 'latest'.")
+            elif conformer_rel_pos_type == "latest":
+                assert conformer_pos_enc_layer_type != "legacy_rel_pos"
+                assert conformer_self_attn_layer_type != "legacy_rel_selfattn"
+            else:
+                raise ValueError(
+                    f"Unknown rel_pos_type: {conformer_rel_pos_type}")
+
+        # define encoder
+        encoder_input_layer = nn.Embedding(
+            num_embeddings=idim,
+            embedding_dim=adim,
+            padding_idx=self.padding_idx)
+        if encoder_type == "transformer":
+            self.encoder = TransformerEncoder(
+                idim=idim,
+                attention_dim=adim,
+                attention_heads=aheads,
+                linear_units=eunits,
+                num_blocks=elayers,
+                input_layer=encoder_input_layer,
+                dropout_rate=transformer_enc_dropout_rate,
+                positional_dropout_rate=transformer_enc_positional_dropout_rate,
+                attention_dropout_rate=transformer_enc_attn_dropout_rate,
+                pos_enc_layer_type=transformer_pos_enc_layer_type,
+                normalize_before=encoder_normalize_before,
+                concat_after=encoder_concat_after,
+                positionwise_layer_type=positionwise_layer_type,
+                positionwise_conv_kernel_size=positionwise_conv_kernel_size,
+                activation_type=transformer_activation_type)
+        elif encoder_type == "conformer":
+            self.encoder = ConformerEncoder(
+                idim=idim,
+                attention_dim=adim,
+                attention_heads=aheads,
+                linear_units=eunits,
+                num_blocks=elayers,
+                input_layer=encoder_input_layer,
+                dropout_rate=transformer_enc_dropout_rate,
+                positional_dropout_rate=transformer_enc_positional_dropout_rate,
+                attention_dropout_rate=transformer_enc_attn_dropout_rate,
+                normalize_before=encoder_normalize_before,
+                concat_after=encoder_concat_after,
+                positionwise_layer_type=positionwise_layer_type,
+                positionwise_conv_kernel_size=positionwise_conv_kernel_size,
+                macaron_style=use_macaron_style_in_conformer,
+                pos_enc_layer_type=conformer_pos_enc_layer_type,
+                selfattention_layer_type=conformer_self_attn_layer_type,
+                activation_type=conformer_activation_type,
+                use_cnn_module=use_cnn_in_conformer,
+                cnn_module_kernel=conformer_enc_kernel_size,
+                zero_triu=zero_triu, )
+        else:
+            raise ValueError(f"{encoder_type} is not supported.")
+
+        # define GST
+        if self.use_gst:
+            self.gst = StyleEncoder(
+                idim=odim,  # the input is mel-spectrogram
+                gst_tokens=gst_tokens,
+                gst_token_dim=adim,
+                gst_heads=gst_heads,
+                conv_layers=gst_conv_layers,
+                conv_chans_list=gst_conv_chans_list,
+                conv_kernel_size=gst_conv_kernel_size,
+                conv_stride=gst_conv_stride,
+                gru_layers=gst_gru_layers,
+                gru_units=gst_gru_units, )
+
+        # define spk and lang embedding
+        self.spks = None
+        if spks is not None and spks > 1:
+            self.spks = spks
+            self.sid_emb = nn.Embedding(spks, adim)
+        self.langs = None
+        if langs is not None and langs > 1:
+            self.langs = langs
+            self.lid_emb = nn.Embedding(langs, adim)
+
+        # define additional projection for speaker embedding
+        self.spk_embed_dim = None
+        if spk_embed_dim is not None and spk_embed_dim > 0:
+            self.spk_embed_dim = spk_embed_dim
+            self.spk_embed_integration_type = spk_embed_integration_type
+        if self.spk_embed_dim is not None:
+            if self.spk_embed_integration_type == "add":
+                self.projection = nn.Linear(self.spk_embed_dim, adim)
+            else:
+                self.projection = nn.Linear(adim + self.spk_embed_dim, adim)
+
+        # define duration predictor
+        self.duration_predictor = DurationPredictor(
+            idim=adim,
+            n_layers=duration_predictor_layers,
+            n_chans=duration_predictor_chans,
+            kernel_size=duration_predictor_kernel_size,
+            dropout_rate=duration_predictor_dropout_rate, )
+
+        # define pitch predictor
+        self.pitch_predictor = VariancePredictor(
+            idim=adim,
+            n_layers=pitch_predictor_layers,
+            n_chans=pitch_predictor_chans,
+            kernel_size=pitch_predictor_kernel_size,
+            dropout_rate=pitch_predictor_dropout, )
+        # NOTE(kan-bayashi): We use continuous pitch + FastPitch style avg
+        self.pitch_embed = nn.Sequential(
+            nn.Conv1D(
+                in_channels=1,
+                out_channels=adim,
+                kernel_size=pitch_embed_kernel_size,
+                padding=(pitch_embed_kernel_size - 1) // 2, ),
+            nn.Dropout(pitch_embed_dropout), )
+
+        # define energy predictor
+        self.energy_predictor = VariancePredictor(
+            idim=adim,
+            n_layers=energy_predictor_layers,
+            n_chans=energy_predictor_chans,
+            kernel_size=energy_predictor_kernel_size,
+            dropout_rate=energy_predictor_dropout, )
+        # NOTE(kan-bayashi): We use continuous enegy + FastPitch style avg
+        self.energy_embed = nn.Sequential(
+            nn.Conv1D(
+                in_channels=1,
+                out_channels=adim,
+                kernel_size=energy_embed_kernel_size,
+                padding=(energy_embed_kernel_size - 1) // 2, ),
+            nn.Dropout(energy_embed_dropout), )
+
+        # define length regulator
+        self.length_regulator = GaussianUpsampling()
+
+        # define decoder
+        # NOTE: we use encoder as decoder
+        # because fastspeech's decoder is the same as encoder
+        if decoder_type == "transformer":
+            self.decoder = TransformerEncoder(
+                idim=0,
+                attention_dim=adim,
+                attention_heads=aheads,
+                linear_units=dunits,
+                num_blocks=dlayers,
+                # in decoder, don't need layer before pos_enc_class (we use embedding here in encoder)
+                input_layer=None,
+                dropout_rate=transformer_dec_dropout_rate,
+                positional_dropout_rate=transformer_dec_positional_dropout_rate,
+                attention_dropout_rate=transformer_dec_attn_dropout_rate,
+                pos_enc_layer_type=transformer_pos_enc_layer_type,
+                normalize_before=decoder_normalize_before,
+                concat_after=decoder_concat_after,
+                positionwise_layer_type=positionwise_layer_type,
+                positionwise_conv_kernel_size=positionwise_conv_kernel_size,
+                activation_type=conformer_activation_type, )
+
+        elif decoder_type == "conformer":
+            self.decoder = ConformerEncoder(
+                idim=0,
+                attention_dim=adim,
+                attention_heads=aheads,
+                linear_units=dunits,
+                num_blocks=dlayers,
+                input_layer=None,
+                dropout_rate=transformer_dec_dropout_rate,
+                positional_dropout_rate=transformer_dec_positional_dropout_rate,
+                attention_dropout_rate=transformer_dec_attn_dropout_rate,
+                normalize_before=decoder_normalize_before,
+                concat_after=decoder_concat_after,
+                positionwise_layer_type=positionwise_layer_type,
+                positionwise_conv_kernel_size=positionwise_conv_kernel_size,
+                macaron_style=use_macaron_style_in_conformer,
+                pos_enc_layer_type=conformer_pos_enc_layer_type,
+                selfattention_layer_type=conformer_self_attn_layer_type,
+                activation_type=conformer_activation_type,
+                use_cnn_module=use_cnn_in_conformer,
+                cnn_module_kernel=conformer_dec_kernel_size, )
+        else:
+            raise ValueError(f"{decoder_type} is not supported.")
+
+        self.generator = HiFiGANGenerator(
+            in_channels=adim,
+            out_channels=generator_out_channels,
+            channels=generator_channels,
+            global_channels=generator_global_channels,
+            kernel_size=generator_kernel_size,
+            upsample_scales=generator_upsample_scales,
+            upsample_kernel_sizes=generator_upsample_kernel_sizes,
+            resblock_kernel_sizes=generator_resblock_kernel_sizes,
+            resblock_dilations=generator_resblock_dilations,
+            use_additional_convs=generator_use_additional_convs,
+            bias=generator_bias,
+            nonlinear_activation=generator_nonlinear_activation,
+            nonlinear_activation_params=generator_nonlinear_activation_params,
+            use_weight_norm=generator_use_weight_norm, )
+
+        self.alignment_module = AlignmentModule(adim, odim)
+
+        # initialize parameters
+        self._reset_parameters(
+            init_type=init_type,
+            init_enc_alpha=init_enc_alpha,
+            init_dec_alpha=init_dec_alpha, )
+
+    def forward(
+            self,
+            text: paddle.Tensor,
+            text_lengths: paddle.Tensor,
+            feats: paddle.Tensor,
+            feats_lengths: paddle.Tensor,
+            durations: paddle.Tensor,
+            durations_lengths: paddle.Tensor,
+            pitch: paddle.Tensor,
+            energy: paddle.Tensor,
+            sids: Optional[paddle.Tensor]=None,
+            spembs: Optional[paddle.Tensor]=None,
+            lids: Optional[paddle.Tensor]=None,
+            use_alignment_module: bool=False,
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor, paddle.Tensor,
+               paddle.Tensor, paddle.Tensor,
+               Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor, paddle.Tensor,
+                     paddle.Tensor, paddle.Tensor, ], ]:
+        """Calculate forward propagation.
+        Args:
+            text (Tensor):
+                Text index tensor (B, T_text).
+            text_lengths (Tensor):
+                Text length tensor (B,).
+            feats (Tensor):
+                Feature tensor (B, aux_channels, T_feats).
+            feats_lengths (Tensor):
+                Feature length tensor (B,).
+            pitch (Tensor): 
+                Batch of padded token-averaged pitch (B, T_text, 1).
+            energy (Tensor):
+                Batch of padded token-averaged energy (B, T_text, 1).
+            sids (Optional[Tensor]):
+                Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]):
+                Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]):
+                Language index tensor (B,) or (B, 1).
+            use_alignment_module (bool):
+                Whether to use alignment module.
+                
+        Returns:
+            Tensor: 
+                Waveform tensor (B, 1, segment_size * upsample_factor).
+            Tensor: 
+                binarization loss ()
+            Tensor: 
+                log probability attention matrix (B,T_feats,T_text)
+            Tensor: 
+                Segments start index tensor (B,).
+            Tensor: 
+                predicted duration (B,T_text)
+            Tensor: 
+                ground-truth duration obtained from an alignment module (B,T_text)
+            Tensor: 
+                predicted pitch (B,T_text,1)
+            Tensor: 
+                ground-truth averaged pitch (B,T_text,1)
+            Tensor: 
+                predicted energy (B,T_text,1)
+            Tensor: 
+                ground-truth averaged energy (B,T_text,1)
+        """
+        if use_alignment_module:
+            text = text[:, :text_lengths.max()]  # for data-parallel
+            feats = feats[:, :feats_lengths.max()]  # for data-parallel
+            pitch = pitch[:, :durations_lengths.max()]  # for data-parallel
+            energy = energy[:, :durations_lengths.max()]  # for data-parallel
+        else:
+            text = text[:, :text_lengths.max()]  # for data-parallel
+            feats = feats[:, :feats_lengths.max()]  # for data-parallel
+            pitch = pitch[:, :feats_lengths.max()]  # for data-parallel
+            energy = energy[:, :feats_lengths.max()]  # for data-parallel
+
+        # forward encoder
+        x_masks = self._source_mask(text_lengths)
+        hs, _ = self.encoder(text, x_masks)  # (B, T_text, adim)
+
+        # integrate with GST
+        if self.use_gst:
+            style_embs = self.gst(ys)
+            hs = hs + style_embs.unsqueeze(1)
+
+        # integrate with SID and LID embeddings
+        if self.spks is not None:
+            sid_embs = self.sid_emb(sids.reshape([-1]))
+            hs = hs + sid_embs.unsqueeze(1)
+        if self.langs is not None:
+            lid_embs = self.lid_emb(lids.reshape([-1]))
+            hs = hs + lid_embs.unsqueeze(1)
+
+        # integrate speaker embedding
+        if self.spk_embed_dim is not None:
+            hs = self._integrate_with_spk_embed(hs, spembs)
+
+        # forward alignment module and obtain duration, averaged pitch, energy
+        h_masks = make_pad_mask(text_lengths)
+        if use_alignment_module:
+            log_p_attn = self.alignment_module(hs, feats, h_masks)
+            ds, bin_loss = viterbi_decode(log_p_attn, text_lengths,
+                                          feats_lengths)
+            ps = average_by_duration(ds,
+                                     pitch.squeeze(-1), text_lengths,
+                                     feats_lengths).unsqueeze(-1)
+            es = average_by_duration(ds,
+                                     energy.squeeze(-1), text_lengths,
+                                     feats_lengths).unsqueeze(-1)
+        else:
+            ds = durations
+            ps = pitch
+            es = energy
+            log_p_attn = attn = bin_loss = None
+
+        # forward duration predictor and variance predictors
+        if self.stop_gradient_from_pitch_predictor:
+            p_outs = self.pitch_predictor(hs.detach(), h_masks.unsqueeze(-1))
+        else:
+            p_outs = self.pitch_predictor(hs, h_masks.unsqueeze(-1))
+        if self.stop_gradient_from_energy_predictor:
+            e_outs = self.energy_predictor(hs.detach(), h_masks.unsqueeze(-1))
+        else:
+            e_outs = self.energy_predictor(hs, h_masks.unsqueeze(-1))
+
+        d_outs = self.duration_predictor(hs, h_masks)
+
+        # use groundtruth in training
+        p_embs = self.pitch_embed(ps.transpose([0, 2, 1])).transpose([0, 2, 1])
+        e_embs = self.energy_embed(es.transpose([0, 2, 1])).transpose([0, 2, 1])
+        hs = hs + e_embs + p_embs
+
+        # upsampling
+        h_masks = make_non_pad_mask(feats_lengths)
+        # d_masks = make_non_pad_mask(text_lengths).to(ds.device)
+        d_masks = make_non_pad_mask(text_lengths)
+        hs = self.length_regulator(hs, ds, h_masks,
+                                   d_masks)  # (B, T_feats, adim)
+
+        # forward decoder
+        h_masks = self._source_mask(feats_lengths)
+        zs, _ = self.decoder(hs, h_masks)  # (B, T_feats, adim)
+
+        # get random segments
+        z_segments, z_start_idxs = get_random_segments(
+            zs.transpose([0, 2, 1]),
+            feats_lengths,
+            self.segment_size, )
+        # forward generator
+        wav = self.generator(z_segments)
+        if use_alignment_module:
+            return wav, bin_loss, log_p_attn, z_start_idxs, d_outs, ds, p_outs, ps, e_outs, es
+        else:
+            return wav, None, None, z_start_idxs, d_outs, ds, p_outs, ps, e_outs, es
+
+    def inference(
+            self,
+            text: paddle.Tensor,
+            text_lengths: paddle.Tensor,
+            feats: Optional[paddle.Tensor]=None,
+            feats_lengths: Optional[paddle.Tensor]=None,
+            pitch: Optional[paddle.Tensor]=None,
+            energy: Optional[paddle.Tensor]=None,
+            sids: Optional[paddle.Tensor]=None,
+            spembs: Optional[paddle.Tensor]=None,
+            lids: Optional[paddle.Tensor]=None,
+            use_alignment_module: bool=False,
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Run inference.
+
+        Args:
+            text (Tensor): Input text index tensor (B, T_text,).
+            text_lengths (Tensor): Text length tensor (B,).
+            feats (Tensor): Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor): Feature length tensor (B,).
+            pitch (Tensor): Pitch tensor (B, T_feats, 1)
+            energy (Tensor): Energy tensor (B, T_feats, 1)
+            sids (Optional[Tensor]): Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]): Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]): Language index tensor (B,) or (B, 1).
+            use_alignment_module (bool): Whether to use alignment module.
+
+        Returns:
+            Tensor: Generated waveform tensor (B, T_wav).
+            Tensor: Duration tensor (B, T_text).
+
+        """
+        # forward encoder
+        x_masks = self._source_mask(text_lengths)
+        hs, _ = self.encoder(text, x_masks)  # (B, T_text, adim)
+
+        # integrate with GST
+        if self.use_gst:
+            style_embs = self.gst(ys)
+            hs = hs + style_embs.unsqueeze(1)
+
+        # integrate with SID and LID embeddings
+        if self.spks is not None:
+            sid_embs = self.sid_emb(sids.view(-1))
+            hs = hs + sid_embs.unsqueeze(1)
+        if self.langs is not None:
+            lid_embs = self.lid_emb(lids.view(-1))
+            hs = hs + lid_embs.unsqueeze(1)
+
+        # integrate speaker embedding
+        if self.spk_embed_dim is not None:
+            hs = self._integrate_with_spk_embed(hs, spembs)
+
+        h_masks = make_pad_mask(text_lengths)
+        if use_alignment_module:
+            # forward alignment module and obtain duration, averaged pitch, energy
+            log_p_attn, attn = self.alignment_module(hs, feats, h_masks)
+            d_outs, _ = viterbi_decode(log_p_attn, text_lengths, feats_lengths)
+            p_outs = average_by_duration(d_outs,
+                                         pitch.squeeze(-1), text_lengths,
+                                         feats_lengths).unsqueeze(-1)
+            e_outs = average_by_duration(d_outs,
+                                         energy.squeeze(-1), text_lengths,
+                                         feats_lengths).unsqueeze(-1)
+        else:
+            # forward duration predictor and variance predictors
+            p_outs = self.pitch_predictor(hs, h_masks.unsqueeze(-1))
+            e_outs = self.energy_predictor(hs, h_masks.unsqueeze(-1))
+            d_outs = self.duration_predictor.inference(hs, h_masks)
+
+        p_embs = self.pitch_embed(p_outs.transpose([0, 2, 1])).transpose(
+            [0, 2, 1])
+        e_embs = self.energy_embed(e_outs.transpose([0, 2, 1])).transpose(
+            [0, 2, 1])
+        hs = hs + e_embs + p_embs
+
+        # upsampling
+        if feats_lengths is not None:
+            h_masks = make_non_pad_mask(feats_lengths)
+        else:
+            h_masks = None
+        d_masks = make_non_pad_mask(text_lengths)
+        hs = self.length_regulator(hs, d_outs, h_masks,
+                                   d_masks)  # (B, T_feats, adim)
+
+        # forward decoder
+        if feats_lengths is not None:
+            h_masks = self._source_mask(feats_lengths)
+        else:
+            h_masks = None
+        zs, _ = self.decoder(hs, h_masks)  # (B, T_feats, adim)
+
+        # forward generator
+        wav = self.generator(zs.transpose([0, 2, 1]))
+
+        return wav.squeeze(1), d_outs
+
+    def _integrate_with_spk_embed(self,
+                                  hs: paddle.Tensor,
+                                  spembs: paddle.Tensor) -> paddle.Tensor:
+        """Integrate speaker embedding with hidden states.
+
+        Args:
+            hs (Tensor): Batch of hidden state sequences (B, T_text, adim).
+            spembs (Tensor): Batch of speaker embeddings (B, spk_embed_dim).
+
+        Returns:
+            Tensor: Batch of integrated hidden state sequences (B, T_text, adim).
+
+        """
+        if self.spk_embed_integration_type == "add":
+            # apply projection and then add to hidden states
+            spembs = self.projection(F.normalize(spembs))
+            hs = hs + spembs.unsqueeze(1)
+        elif self.spk_embed_integration_type == "concat":
+            # concat hidden states with spk embeds and then apply projection
+            spembs = F.normalize(spembs).unsqueeze(1).expand(-1, hs.shape[1],
+                                                             -1)
+            hs = self.projection(paddle.concat([hs, spembs], axis=-1))
+        else:
+            raise NotImplementedError("support only add or concat.")
+
+        return hs
+
+    def _generate_path(self, dur: paddle.Tensor,
+                       mask: paddle.Tensor) -> paddle.Tensor:
+        """Generate path a.k.a. monotonic attention.
+        Args:
+            dur (Tensor):
+                Duration tensor (B, 1, T_text).
+            mask (Tensor):
+                Attention mask tensor (B, 1, T_feats, T_text).
+        Returns:
+            Tensor:
+                Path tensor (B, 1, T_feats, T_text).
+        """
+        b, _, t_y, t_x = paddle.shape(mask)
+        cum_dur = paddle.cumsum(dur, -1)
+        cum_dur_flat = paddle.reshape(cum_dur, [b * t_x])
+
+        path = paddle.arange(t_y, dtype=dur.dtype)
+        path = path.unsqueeze(0) < cum_dur_flat.unsqueeze(1)
+        path = paddle.reshape(path, [b, t_x, t_y])
+        '''
+        path will be like (t_x = 3, t_y = 5):
+        [[[1., 1., 0., 0., 0.],      [[[1., 1., 0., 0., 0.],
+          [1., 1., 1., 1., 0.],  -->   [0., 0., 1., 1., 0.],
+          [1., 1., 1., 1., 1.]]]       [0., 0., 0., 0., 1.]]]
+        '''
+
+        path = paddle.cast(path, dtype='float32')
+        pad_tmp = self.pad1d(path)[:, :-1]
+        path = path - pad_tmp
+        return path.unsqueeze(1).transpose([0, 1, 3, 2]) * mask
+
+    def _source_mask(self, ilens: paddle.Tensor) -> paddle.Tensor:
+        """Make masks for self-attention.
+
+        Args:
+            ilens (LongTensor): Batch of lengths (B,).
+
+        Returns:
+            Tensor: Mask tensor for self-attention.
+                dtype=paddle.uint8 
+
+        Examples:
+            >>> ilens = [5, 3]
+            >>> self._source_mask(ilens)
+            tensor([[[1, 1, 1, 1, 1],
+                     [1, 1, 1, 0, 0]]], dtype=torch.uint8)
+
+        """
+        x_masks = paddle.to_tensor(make_non_pad_mask(ilens))
+        return x_masks.unsqueeze(-2)
+
+    def _reset_parameters(self,
+                          init_type: str,
+                          init_enc_alpha: float,
+                          init_dec_alpha: float):
+        # initialize parameters
+        initialize(self, init_type)
+
+        # initialize alpha in scaled positional encoding
+        if self.encoder_type == "transformer" and self.use_scaled_pos_enc:
+            self.encoder.embed[-1].alpha.data = paddle.to_tensor(init_enc_alpha)
+        if self.decoder_type == "transformer" and self.use_scaled_pos_enc:
+            self.decoder.embed[-1].alpha.data = paddle.to_tensor(init_dec_alpha)

paddlespeech/t2s/models/jets/jets.py

@@ -0,0 +1,582 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Generator module in JETS.
+
+This code is based on https://github.com/imdanboy/jets.
+
+"""
+"""JETS module"""
+import math
+from typing import Any
+from typing import Dict
+from typing import Optional
+
+import paddle
+from paddle import nn
+from typeguard import check_argument_types
+
+from paddlespeech.t2s.models.hifigan import HiFiGANMultiPeriodDiscriminator
+from paddlespeech.t2s.models.hifigan import HiFiGANMultiScaleDiscriminator
+from paddlespeech.t2s.models.hifigan import HiFiGANMultiScaleMultiPeriodDiscriminator
+from paddlespeech.t2s.models.hifigan import HiFiGANPeriodDiscriminator
+from paddlespeech.t2s.models.hifigan import HiFiGANScaleDiscriminator
+from paddlespeech.t2s.models.jets.generator import JETSGenerator
+from paddlespeech.utils.initialize import _calculate_fan_in_and_fan_out
+from paddlespeech.utils.initialize import kaiming_uniform_
+from paddlespeech.utils.initialize import normal_
+from paddlespeech.utils.initialize import ones_
+from paddlespeech.utils.initialize import uniform_
+from paddlespeech.utils.initialize import zeros_
+
+AVAILABLE_GENERATERS = {
+    "jets_generator": JETSGenerator,
+}
+AVAILABLE_DISCRIMINATORS = {
+    "hifigan_period_discriminator":
+    HiFiGANPeriodDiscriminator,
+    "hifigan_scale_discriminator":
+    HiFiGANScaleDiscriminator,
+    "hifigan_multi_period_discriminator":
+    HiFiGANMultiPeriodDiscriminator,
+    "hifigan_multi_scale_discriminator":
+    HiFiGANMultiScaleDiscriminator,
+    "hifigan_multi_scale_multi_period_discriminator":
+    HiFiGANMultiScaleMultiPeriodDiscriminator,
+}
+
+
+class JETS(nn.Layer):
+    """JETS module (generator + discriminator).
+    This is a module of JETS described in `JETS: Jointly Training FastSpeech2 
+    and HiFi-GAN for End to End Text to Speech`_.
+    .. _`JETS: Jointly Training FastSpeech2 and HiFi-GAN for End to End Text to Speech
+        Text-to-Speech`: https://arxiv.org/abs/2203.16852v1
+    """
+
+    def __init__(
+            self,
+            # generator related
+            idim: int,
+            odim: int,
+            sampling_rate: int=22050,
+            generator_type: str="jets_generator",
+            generator_params: Dict[str, Any]={
+                "adim": 256,
+                "aheads": 2,
+                "elayers": 4,
+                "eunits": 1024,
+                "dlayers": 4,
+                "dunits": 1024,
+                "positionwise_layer_type": "conv1d",
+                "positionwise_conv_kernel_size": 1,
+                "use_scaled_pos_enc": True,
+                "use_batch_norm": True,
+                "encoder_normalize_before": True,
+                "decoder_normalize_before": True,
+                "encoder_concat_after": False,
+                "decoder_concat_after": False,
+                "reduction_factor": 1,
+                "encoder_type": "transformer",
+                "decoder_type": "transformer",
+                "transformer_enc_dropout_rate": 0.1,
+                "transformer_enc_positional_dropout_rate": 0.1,
+                "transformer_enc_attn_dropout_rate": 0.1,
+                "transformer_dec_dropout_rate": 0.1,
+                "transformer_dec_positional_dropout_rate": 0.1,
+                "transformer_dec_attn_dropout_rate": 0.1,
+                "conformer_rel_pos_type": "latest",
+                "conformer_pos_enc_layer_type": "rel_pos",
+                "conformer_self_attn_layer_type": "rel_selfattn",
+                "conformer_activation_type": "swish",
+                "use_macaron_style_in_conformer": True,
+                "use_cnn_in_conformer": True,
+                "zero_triu": False,
+                "conformer_enc_kernel_size": 7,
+                "conformer_dec_kernel_size": 31,
+                "duration_predictor_layers": 2,
+                "duration_predictor_chans": 384,
+                "duration_predictor_kernel_size": 3,
+                "duration_predictor_dropout_rate": 0.1,
+                "energy_predictor_layers": 2,
+                "energy_predictor_chans": 384,
+                "energy_predictor_kernel_size": 3,
+                "energy_predictor_dropout": 0.5,
+                "energy_embed_kernel_size": 1,
+                "energy_embed_dropout": 0.5,
+                "stop_gradient_from_energy_predictor": False,
+                "pitch_predictor_layers": 5,
+                "pitch_predictor_chans": 384,
+                "pitch_predictor_kernel_size": 5,
+                "pitch_predictor_dropout": 0.5,
+                "pitch_embed_kernel_size": 1,
+                "pitch_embed_dropout": 0.5,
+                "stop_gradient_from_pitch_predictor": True,
+                "generator_out_channels": 1,
+                "generator_channels": 512,
+                "generator_global_channels": -1,
+                "generator_kernel_size": 7,
+                "generator_upsample_scales": [8, 8, 2, 2],
+                "generator_upsample_kernel_sizes": [16, 16, 4, 4],
+                "generator_resblock_kernel_sizes": [3, 7, 11],
+                "generator_resblock_dilations":
+                [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                "generator_use_additional_convs": True,
+                "generator_bias": True,
+                "generator_nonlinear_activation": "LeakyReLU",
+                "generator_nonlinear_activation_params": {
+                    "negative_slope": 0.1
+                },
+                "generator_use_weight_norm": True,
+                "segment_size": 64,
+                "spks": -1,
+                "langs": -1,
+                "spk_embed_dim": None,
+                "spk_embed_integration_type": "add",
+                "use_gst": False,
+                "gst_tokens": 10,
+                "gst_heads": 4,
+                "gst_conv_layers": 6,
+                "gst_conv_chans_list": [32, 32, 64, 64, 128, 128],
+                "gst_conv_kernel_size": 3,
+                "gst_conv_stride": 2,
+                "gst_gru_layers": 1,
+                "gst_gru_units": 128,
+                "init_type": "xavier_uniform",
+                "init_enc_alpha": 1.0,
+                "init_dec_alpha": 1.0,
+                "use_masking": False,
+                "use_weighted_masking": False,
+            },
+            # discriminator related
+            discriminator_type: str="hifigan_multi_scale_multi_period_discriminator",
+            discriminator_params: Dict[str, Any]={
+                "scales": 1,
+                "scale_downsample_pooling": "AvgPool1D",
+                "scale_downsample_pooling_params": {
+                    "kernel_size": 4,
+                    "stride": 2,
+                    "padding": 2,
+                },
+                "scale_discriminator_params": {
+                    "in_channels": 1,
+                    "out_channels": 1,
+                    "kernel_sizes": [15, 41, 5, 3],
+                    "channels": 128,
+                    "max_downsample_channels": 1024,
+                    "max_groups": 16,
+                    "bias": True,
+                    "downsample_scales": [2, 2, 4, 4, 1],
+                    "nonlinear_activation": "leakyrelu",
+                    "nonlinear_activation_params": {
+                        "negative_slope": 0.1
+                    },
+                    "use_weight_norm": True,
+                    "use_spectral_norm": False,
+                },
+                "follow_official_norm": False,
+                "periods": [2, 3, 5, 7, 11],
+                "period_discriminator_params": {
+                    "in_channels": 1,
+                    "out_channels": 1,
+                    "kernel_sizes": [5, 3],
+                    "channels": 32,
+                    "downsample_scales": [3, 3, 3, 3, 1],
+                    "max_downsample_channels": 1024,
+                    "bias": True,
+                    "nonlinear_activation": "leakyrelu",
+                    "nonlinear_activation_params": {
+                        "negative_slope": 0.1
+                    },
+                    "use_weight_norm": True,
+                    "use_spectral_norm": False,
+                },
+            },
+            cache_generator_outputs: bool=True, ):
+        """Initialize JETS module.
+        Args:
+            idim (int):
+                Input vocabrary size.
+            odim (int):
+                Acoustic feature dimension. The actual output channels will
+                be 1 since JETS is the end-to-end text-to-wave model but for the
+                compatibility odim is used to indicate the acoustic feature dimension.
+            sampling_rate (int):
+                Sampling rate, not used for the training but it will
+                be referred in saving waveform during the inference.
+            generator_type (str):
+                Generator type.
+            generator_params (Dict[str, Any]):
+                Parameter dict for generator.
+            discriminator_type (str):
+                Discriminator type.
+            discriminator_params (Dict[str, Any]):
+                Parameter dict for discriminator.
+            cache_generator_outputs (bool):
+                Whether to cache generator outputs.
+        """
+        assert check_argument_types()
+        super().__init__()
+
+        # define modules
+        generator_class = AVAILABLE_GENERATERS[generator_type]
+        if generator_type == "jets_generator":
+            # NOTE: Update parameters for the compatibility.
+            #   The idim and odim is automatically decided from input data,
+            #   where idim represents #vocabularies and odim represents
+            #   the input acoustic feature dimension.
+            generator_params.update(idim=idim, odim=odim)
+        self.generator = generator_class(
+            **generator_params, )
+        discriminator_class = AVAILABLE_DISCRIMINATORS[discriminator_type]
+        self.discriminator = discriminator_class(
+            **discriminator_params, )
+
+        # cache
+        self.cache_generator_outputs = cache_generator_outputs
+        self._cache = None
+
+        # store sampling rate for saving wav file
+        # (not used for the training)
+        self.fs = sampling_rate
+
+        # store parameters for test compatibility
+        self.spks = self.generator.spks
+        self.langs = self.generator.langs
+        self.spk_embed_dim = self.generator.spk_embed_dim
+
+        self.reuse_cache_gen = True
+        self.reuse_cache_dis = True
+
+        self.reset_parameters()
+        self.generator._reset_parameters(
+            init_type=generator_params["init_type"],
+            init_enc_alpha=generator_params["init_enc_alpha"],
+            init_dec_alpha=generator_params["init_dec_alpha"], )
+
+    def forward(
+            self,
+            text: paddle.Tensor,
+            text_lengths: paddle.Tensor,
+            feats: paddle.Tensor,
+            feats_lengths: paddle.Tensor,
+            durations: paddle.Tensor,
+            durations_lengths: paddle.Tensor,
+            pitch: paddle.Tensor,
+            energy: paddle.Tensor,
+            sids: Optional[paddle.Tensor]=None,
+            spembs: Optional[paddle.Tensor]=None,
+            lids: Optional[paddle.Tensor]=None,
+            forward_generator: bool=True,
+            use_alignment_module: bool=False,
+            **kwargs,
+    ) -> Dict[str, Any]:
+        """Perform generator forward.
+        Args:
+            text (Tensor):
+                Text index tensor (B, T_text).
+            text_lengths (Tensor):
+                Text length tensor (B,).
+            feats (Tensor):
+                Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor):
+                Feature length tensor (B,).
+            durations(Tensor(int64)): 
+                Batch of padded durations (B, Tmax).
+            durations_lengths (Tensor):
+                durations length tensor (B,).
+            pitch(Tensor): 
+                Batch of padded token-averaged pitch (B, Tmax, 1).
+            energy(Tensor): 
+                Batch of padded token-averaged energy (B, Tmax, 1).
+            sids (Optional[Tensor]):
+                Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]):
+                Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]):
+                Language index tensor (B,) or (B, 1).
+            forward_generator (bool):
+                Whether to forward generator.
+            use_alignment_module (bool):
+                Whether to use alignment module.
+        Returns:
+
+        """
+        if forward_generator:
+            return self._forward_generator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                durations=durations,
+                durations_lengths=durations_lengths,
+                pitch=pitch,
+                energy=energy,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+                use_alignment_module=use_alignment_module, )
+        else:
+            return self._forward_discrminator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                durations=durations,
+                durations_lengths=durations_lengths,
+                pitch=pitch,
+                energy=energy,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+                use_alignment_module=use_alignment_module, )
+
+    def _forward_generator(
+            self,
+            text: paddle.Tensor,
+            text_lengths: paddle.Tensor,
+            feats: paddle.Tensor,
+            feats_lengths: paddle.Tensor,
+            durations: paddle.Tensor,
+            durations_lengths: paddle.Tensor,
+            pitch: paddle.Tensor,
+            energy: paddle.Tensor,
+            sids: Optional[paddle.Tensor]=None,
+            spembs: Optional[paddle.Tensor]=None,
+            lids: Optional[paddle.Tensor]=None,
+            use_alignment_module: bool=False,
+            **kwargs, ) -> Dict[str, Any]:
+        """Perform generator forward.
+        Args:
+            text (Tensor):
+                Text index tensor (B, T_text).
+            text_lengths (Tensor):
+                Text length tensor (B,).
+            feats (Tensor):
+                Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor):
+                Feature length tensor (B,).
+            durations(Tensor(int64)): 
+                Batch of padded durations (B, Tmax).
+            durations_lengths (Tensor):
+                durations length tensor (B,).
+            pitch(Tensor): 
+                Batch of padded token-averaged pitch (B, Tmax, 1).
+            energy(Tensor): 
+                Batch of padded token-averaged energy (B, Tmax, 1).
+            sids (Optional[Tensor]):
+                Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]):
+                Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]):
+                Language index tensor (B,) or (B, 1).
+            use_alignment_module (bool):
+                Whether to use alignment module.
+        Returns:
+
+        """
+        # setup
+        # calculate generator outputs
+        self.reuse_cache_gen = True
+        if not self.cache_generator_outputs or self._cache is None:
+            self.reuse_cache_gen = False
+            outs = self.generator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                durations=durations,
+                durations_lengths=durations_lengths,
+                pitch=pitch,
+                energy=energy,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+                use_alignment_module=use_alignment_module, )
+        else:
+            outs = self._cache
+
+        # store cache
+        if self.training and self.cache_generator_outputs and not self.reuse_cache_gen:
+            self._cache = outs
+
+        return outs
+
+    def _forward_discrminator(
+            self,
+            text: paddle.Tensor,
+            text_lengths: paddle.Tensor,
+            feats: paddle.Tensor,
+            feats_lengths: paddle.Tensor,
+            durations: paddle.Tensor,
+            durations_lengths: paddle.Tensor,
+            pitch: paddle.Tensor,
+            energy: paddle.Tensor,
+            sids: Optional[paddle.Tensor]=None,
+            spembs: Optional[paddle.Tensor]=None,
+            lids: Optional[paddle.Tensor]=None,
+            use_alignment_module: bool=False,
+            **kwargs, ) -> Dict[str, Any]:
+        """Perform discriminator forward.
+        Args:
+            text (Tensor):
+                Text index tensor (B, T_text).
+            text_lengths (Tensor):
+                Text length tensor (B,).
+            feats (Tensor):
+                Feature tensor (B, T_feats, aux_channels).
+            feats_lengths (Tensor):
+                Feature length tensor (B,).
+            durations(Tensor(int64)): 
+                Batch of padded durations (B, Tmax).
+            durations_lengths (Tensor):
+                durations length tensor (B,).
+            pitch(Tensor): 
+                Batch of padded token-averaged pitch (B, Tmax, 1).
+            energy(Tensor): 
+                Batch of padded token-averaged energy (B, Tmax, 1).
+            sids (Optional[Tensor]):
+                Speaker index tensor (B,) or (B, 1).
+            spembs (Optional[Tensor]):
+                Speaker embedding tensor (B, spk_embed_dim).
+            lids (Optional[Tensor]):
+                Language index tensor (B,) or (B, 1).
+            use_alignment_module (bool):
+                Whether to use alignment module.
+        Returns:
+
+        """
+        # setup
+        # calculate generator outputs
+        self.reuse_cache_dis = True
+        if not self.cache_generator_outputs or self._cache is None:
+            self.reuse_cache_dis = False
+            outs = self.generator(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                durations=durations,
+                durations_lengths=durations_lengths,
+                pitch=pitch,
+                energy=energy,
+                sids=sids,
+                spembs=spembs,
+                lids=lids,
+                use_alignment_module=use_alignment_module,
+                **kwargs, )
+        else:
+            outs = self._cache
+
+        # store cache
+        if self.cache_generator_outputs and not self.reuse_cache_dis:
+            self._cache = outs
+
+        return outs
+
+    def inference(self,
+                  text: paddle.Tensor,
+                  feats: Optional[paddle.Tensor]=None,
+                  pitch: Optional[paddle.Tensor]=None,
+                  energy: Optional[paddle.Tensor]=None,
+                  use_alignment_module: bool=False,
+                  **kwargs) -> Dict[str, paddle.Tensor]:
+        """Run inference.
+        Args:
+            text (Tensor):
+                Input text index tensor (T_text,).
+            feats (Tensor):
+                Feature tensor (T_feats, aux_channels).
+            pitch (Tensor):
+                Pitch tensor (T_feats, 1).
+            energy (Tensor): 
+                Energy tensor (T_feats, 1).
+            use_alignment_module (bool):
+                Whether to use alignment module.
+        Returns:
+            Dict[str, Tensor]:
+                * wav (Tensor):
+                    Generated waveform tensor (T_wav,).
+                * duration (Tensor):
+                    Predicted duration tensor (T_text,).
+        """
+        # setup
+        text = text[None]
+        text_lengths = paddle.to_tensor(paddle.shape(text)[1])
+
+        # inference
+        if use_alignment_module:
+            assert feats is not None
+            feats = feats[None]
+            feats_lengths = paddle.to_tensor(paddle.shape(feats)[1])
+            pitch = pitch[None]
+            energy = energy[None]
+            wav, dur = self.generator.inference(
+                text=text,
+                text_lengths=text_lengths,
+                feats=feats,
+                feats_lengths=feats_lengths,
+                pitch=pitch,
+                energy=energy,
+                use_alignment_module=use_alignment_module,
+                **kwargs)
+        else:
+            wav, dur = self.generator.inference(
+                text=text,
+                text_lengths=text_lengths,
+                **kwargs, )
+        return dict(wav=paddle.reshape(wav, [-1]), duration=dur[0])
+
+    def reset_parameters(self):
+        def _reset_parameters(module):
+            if isinstance(
+                    module,
+                (nn.Conv1D, nn.Conv1DTranspose, nn.Conv2D, nn.Conv2DTranspose)):
+                kaiming_uniform_(module.weight, a=math.sqrt(5))
+                if module.bias is not None:
+                    fan_in, _ = _calculate_fan_in_and_fan_out(module.weight)
+                    if fan_in != 0:
+                        bound = 1 / math.sqrt(fan_in)
+                        uniform_(module.bias, -bound, bound)
+
+            if isinstance(
+                    module,
+                (nn.BatchNorm1D, nn.BatchNorm2D, nn.GroupNorm, nn.LayerNorm)):
+                ones_(module.weight)
+                zeros_(module.bias)
+
+            if isinstance(module, nn.Linear):
+                kaiming_uniform_(module.weight, a=math.sqrt(5))
+                if module.bias is not None:
+                    fan_in, _ = _calculate_fan_in_and_fan_out(module.weight)
+                    bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
+                    uniform_(module.bias, -bound, bound)
+
+            if isinstance(module, nn.Embedding):
+                normal_(module.weight)
+                if module._padding_idx is not None:
+                    with paddle.no_grad():
+                        module.weight[module._padding_idx] = 0
+
+        self.apply(_reset_parameters)
+
+
+class JETSInference(nn.Layer):
+    def __init__(self, model):
+        super().__init__()
+        self.acoustic_model = model
+
+    def forward(self, text, sids=None):
+        out = self.acoustic_model.inference(text)
+        wav = out['wav']
+        return wav

paddlespeech/t2s/models/jets/jets_updater.py

@@ -0,0 +1,437 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Generator module in JETS.
+
+This code is based on https://github.com/imdanboy/jets.
+
+"""
+import logging
+from typing import Dict
+
+import paddle
+from paddle import distributed as dist
+from paddle.io import DataLoader
+from paddle.nn import Layer
+from paddle.optimizer import Optimizer
+from paddle.optimizer.lr import LRScheduler
+
+from paddlespeech.t2s.modules.nets_utils import get_segments
+from paddlespeech.t2s.training.extensions.evaluator import StandardEvaluator
+from paddlespeech.t2s.training.reporter import report
+from paddlespeech.t2s.training.updaters.standard_updater import StandardUpdater
+from paddlespeech.t2s.training.updaters.standard_updater import UpdaterState
+
+logging.basicConfig(
+    format='%(asctime)s [%(levelname)s] [%(filename)s:%(lineno)d] %(message)s',
+    datefmt='[%Y-%m-%d %H:%M:%S]')
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+
+
+class JETSUpdater(StandardUpdater):
+    def __init__(self,
+                 model: Layer,
+                 optimizers: Dict[str, Optimizer],
+                 criterions: Dict[str, Layer],
+                 schedulers: Dict[str, LRScheduler],
+                 dataloader: DataLoader,
+                 generator_train_start_steps: int=0,
+                 discriminator_train_start_steps: int=100000,
+                 lambda_adv: float=1.0,
+                 lambda_mel: float=45.0,
+                 lambda_feat_match: float=2.0,
+                 lambda_var: float=1.0,
+                 lambda_align: float=2.0,
+                 generator_first: bool=False,
+                 use_alignment_module: bool=False,
+                 output_dir=None):
+        # it is designed to hold multiple models
+        # 因为输入的是单模型，但是没有用到父类的 init(), 所以需要重新写这部分
+        models = {"main": model}
+        self.models: Dict[str, Layer] = models
+        # self.model = model
+
+        self.model = model._layers if isinstance(model,
+                                                 paddle.DataParallel) else model
+
+        self.optimizers = optimizers
+        self.optimizer_g: Optimizer = optimizers['generator']
+        self.optimizer_d: Optimizer = optimizers['discriminator']
+
+        self.criterions = criterions
+        self.criterion_mel = criterions['mel']
+        self.criterion_feat_match = criterions['feat_match']
+        self.criterion_gen_adv = criterions["gen_adv"]
+        self.criterion_dis_adv = criterions["dis_adv"]
+        self.criterion_var = criterions["var"]
+        self.criterion_forwardsum = criterions["forwardsum"]
+
+        self.schedulers = schedulers
+        self.scheduler_g = schedulers['generator']
+        self.scheduler_d = schedulers['discriminator']
+
+        self.dataloader = dataloader
+
+        self.generator_train_start_steps = generator_train_start_steps
+        self.discriminator_train_start_steps = discriminator_train_start_steps
+
+        self.lambda_adv = lambda_adv
+        self.lambda_mel = lambda_mel
+        self.lambda_feat_match = lambda_feat_match
+        self.lambda_var = lambda_var
+        self.lambda_align = lambda_align
+
+        self.use_alignment_module = use_alignment_module
+
+        if generator_first:
+            self.turns = ["generator", "discriminator"]
+        else:
+            self.turns = ["discriminator", "generator"]
+
+        self.state = UpdaterState(iteration=0, epoch=0)
+        self.train_iterator = iter(self.dataloader)
+
+        log_file = output_dir / 'worker_{}.log'.format(dist.get_rank())
+        self.filehandler = logging.FileHandler(str(log_file))
+        logger.addHandler(self.filehandler)
+        self.logger = logger
+        self.msg = ""
+
+    def update_core(self, batch):
+        self.msg = "Rank: {}, ".format(dist.get_rank())
+        losses_dict = {}
+
+        for turn in self.turns:
+            speech = batch["speech"]
+            speech = speech.unsqueeze(1)
+            text_lengths = batch["text_lengths"]
+            feats_lengths = batch["feats_lengths"]
+            outs = self.model(
+                text=batch["text"],
+                text_lengths=batch["text_lengths"],
+                feats=batch["feats"],
+                feats_lengths=batch["feats_lengths"],
+                durations=batch["durations"],
+                durations_lengths=batch["durations_lengths"],
+                pitch=batch["pitch"],
+                energy=batch["energy"],
+                sids=batch.get("spk_id", None),
+                spembs=batch.get("spk_emb", None),
+                forward_generator=turn == "generator",
+                use_alignment_module=self.use_alignment_module)
+            # Generator
+            if turn == "generator":
+                # parse outputs
+                speech_hat_, bin_loss, log_p_attn, start_idxs, d_outs, ds, p_outs, ps, e_outs, es = outs
+                speech_ = get_segments(
+                    x=speech,
+                    start_idxs=start_idxs *
+                    self.model.generator.upsample_factor,
+                    segment_size=self.model.generator.segment_size *
+                    self.model.generator.upsample_factor, )
+
+                # calculate discriminator outputs
+                p_hat = self.model.discriminator(speech_hat_)
+                with paddle.no_grad():
+                    # do not store discriminator gradient in generator turn
+                    p = self.model.discriminator(speech_)
+
+                # calculate losses
+                mel_loss = self.criterion_mel(speech_hat_, speech_)
+
+                adv_loss = self.criterion_gen_adv(p_hat)
+                feat_match_loss = self.criterion_feat_match(p_hat, p)
+                dur_loss, pitch_loss, energy_loss = self.criterion_var(
+                    d_outs, ds, p_outs, ps, e_outs, es, text_lengths)
+
+                mel_loss = mel_loss * self.lambda_mel
+                adv_loss = adv_loss * self.lambda_adv
+                feat_match_loss = feat_match_loss * self.lambda_feat_match
+                g_loss = mel_loss + adv_loss + feat_match_loss
+                var_loss = (
+                    dur_loss + pitch_loss + energy_loss) * self.lambda_var
+
+                gen_loss = g_loss + var_loss  #+ align_loss
+
+                report("train/generator_loss", float(gen_loss))
+                report("train/generator_generator_loss", float(g_loss))
+                report("train/generator_variance_loss", float(var_loss))
+                report("train/generator_generator_mel_loss", float(mel_loss))
+                report("train/generator_generator_adv_loss", float(adv_loss))
+                report("train/generator_generator_feat_match_loss",
+                       float(feat_match_loss))
+                report("train/generator_variance_dur_loss", float(dur_loss))
+                report("train/generator_variance_pitch_loss", float(pitch_loss))
+                report("train/generator_variance_energy_loss",
+                       float(energy_loss))
+
+                losses_dict["generator_loss"] = float(gen_loss)
+                losses_dict["generator_generator_loss"] = float(g_loss)
+                losses_dict["generator_variance_loss"] = float(var_loss)
+                losses_dict["generator_generator_mel_loss"] = float(mel_loss)
+                losses_dict["generator_generator_adv_loss"] = float(adv_loss)
+                losses_dict["generator_generator_feat_match_loss"] = float(
+                    feat_match_loss)
+                losses_dict["generator_variance_dur_loss"] = float(dur_loss)
+                losses_dict["generator_variance_pitch_loss"] = float(pitch_loss)
+                losses_dict["generator_variance_energy_loss"] = float(
+                    energy_loss)
+
+                if self.use_alignment_module == True:
+                    forwardsum_loss = self.criterion_forwardsum(
+                        log_p_attn, text_lengths, feats_lengths)
+                    align_loss = (
+                        forwardsum_loss + bin_loss) * self.lambda_align
+                    report("train/generator_alignment_loss", float(align_loss))
+                    report("train/generator_alignment_forwardsum_loss",
+                           float(forwardsum_loss))
+                    report("train/generator_alignment_bin_loss",
+                           float(bin_loss))
+                    losses_dict["generator_alignment_loss"] = float(align_loss)
+                    losses_dict["generator_alignment_forwardsum_loss"] = float(
+                        forwardsum_loss)
+                    losses_dict["generator_alignment_bin_loss"] = float(
+                        bin_loss)
+
+                self.optimizer_g.clear_grad()
+                gen_loss.backward()
+
+                self.optimizer_g.step()
+                self.scheduler_g.step()
+
+                # reset cache
+                if self.model.reuse_cache_gen or not self.model.training:
+                    self.model._cache = None
+
+            # Disctiminator
+            elif turn == "discriminator":
+                # parse outputs
+                speech_hat_, _, _, start_idxs, *_ = outs
+                speech_ = get_segments(
+                    x=speech,
+                    start_idxs=start_idxs *
+                    self.model.generator.upsample_factor,
+                    segment_size=self.model.generator.segment_size *
+                    self.model.generator.upsample_factor, )
+
+                # calculate discriminator outputs
+                p_hat = self.model.discriminator(speech_hat_.detach())
+                p = self.model.discriminator(speech_)
+
+                # calculate losses
+                real_loss, fake_loss = self.criterion_dis_adv(p_hat, p)
+                dis_loss = real_loss + fake_loss
+
+                report("train/real_loss", float(real_loss))
+                report("train/fake_loss", float(fake_loss))
+                report("train/discriminator_loss", float(dis_loss))
+                losses_dict["real_loss"] = float(real_loss)
+                losses_dict["fake_loss"] = float(fake_loss)
+                losses_dict["discriminator_loss"] = float(dis_loss)
+
+                self.optimizer_d.clear_grad()
+                dis_loss.backward()
+
+                self.optimizer_d.step()
+                self.scheduler_d.step()
+
+                # reset cache
+                if self.model.reuse_cache_dis or not self.model.training:
+                    self.model._cache = None
+
+        self.msg += ', '.join('{}: {:>.6f}'.format(k, v)
+                              for k, v in losses_dict.items())
+
+
+class JETSEvaluator(StandardEvaluator):
+    def __init__(self,
+                 model,
+                 criterions: Dict[str, Layer],
+                 dataloader: DataLoader,
+                 lambda_adv: float=1.0,
+                 lambda_mel: float=45.0,
+                 lambda_feat_match: float=2.0,
+                 lambda_var: float=1.0,
+                 lambda_align: float=2.0,
+                 generator_first: bool=False,
+                 use_alignment_module: bool=False,
+                 output_dir=None):
+        # 因为输入的是单模型，但是没有用到父类的 init(), 所以需要重新写这部分
+        models = {"main": model}
+        self.models: Dict[str, Layer] = models
+        # self.model = model
+        self.model = model._layers if isinstance(model,
+                                                 paddle.DataParallel) else model
+
+        self.criterions = criterions
+        self.criterion_mel = criterions['mel']
+        self.criterion_feat_match = criterions['feat_match']
+        self.criterion_gen_adv = criterions["gen_adv"]
+        self.criterion_dis_adv = criterions["dis_adv"]
+        self.criterion_var = criterions["var"]
+        self.criterion_forwardsum = criterions["forwardsum"]
+
+        self.dataloader = dataloader
+
+        self.lambda_adv = lambda_adv
+        self.lambda_mel = lambda_mel
+        self.lambda_feat_match = lambda_feat_match
+        self.lambda_var = lambda_var
+        self.lambda_align = lambda_align
+        self.use_alignment_module = use_alignment_module
+
+        if generator_first:
+            self.turns = ["generator", "discriminator"]
+        else:
+            self.turns = ["discriminator", "generator"]
+
+        log_file = output_dir / 'worker_{}.log'.format(dist.get_rank())
+        self.filehandler = logging.FileHandler(str(log_file))
+        logger.addHandler(self.filehandler)
+        self.logger = logger
+        self.msg = ""
+
+    def evaluate_core(self, batch):
+        # logging.debug("Evaluate: ")
+        self.msg = "Evaluate: "
+        losses_dict = {}
+
+        for turn in self.turns:
+            speech = batch["speech"]
+            speech = speech.unsqueeze(1)
+            text_lengths = batch["text_lengths"]
+            feats_lengths = batch["feats_lengths"]
+            outs = self.model(
+                text=batch["text"],
+                text_lengths=batch["text_lengths"],
+                feats=batch["feats"],
+                feats_lengths=batch["feats_lengths"],
+                durations=batch["durations"],
+                durations_lengths=batch["durations_lengths"],
+                pitch=batch["pitch"],
+                energy=batch["energy"],
+                sids=batch.get("spk_id", None),
+                spembs=batch.get("spk_emb", None),
+                forward_generator=turn == "generator",
+                use_alignment_module=self.use_alignment_module)
+            # Generator
+            if turn == "generator":
+                # parse outputs
+                speech_hat_, bin_loss, log_p_attn, start_idxs, d_outs, ds, p_outs, ps, e_outs, es = outs
+                speech_ = get_segments(
+                    x=speech,
+                    start_idxs=start_idxs *
+                    self.model.generator.upsample_factor,
+                    segment_size=self.model.generator.segment_size *
+                    self.model.generator.upsample_factor, )
+
+                # calculate discriminator outputs
+                p_hat = self.model.discriminator(speech_hat_)
+                with paddle.no_grad():
+                    # do not store discriminator gradient in generator turn
+                    p = self.model.discriminator(speech_)
+
+                # calculate losses
+                mel_loss = self.criterion_mel(speech_hat_, speech_)
+
+                adv_loss = self.criterion_gen_adv(p_hat)
+                feat_match_loss = self.criterion_feat_match(p_hat, p)
+                dur_loss, pitch_loss, energy_loss = self.criterion_var(
+                    d_outs, ds, p_outs, ps, e_outs, es, text_lengths)
+
+                mel_loss = mel_loss * self.lambda_mel
+                adv_loss = adv_loss * self.lambda_adv
+                feat_match_loss = feat_match_loss * self.lambda_feat_match
+                g_loss = mel_loss + adv_loss + feat_match_loss
+                var_loss = (
+                    dur_loss + pitch_loss + energy_loss) * self.lambda_var
+
+                gen_loss = g_loss + var_loss  #+ align_loss
+
+                report("eval/generator_loss", float(gen_loss))
+                report("eval/generator_generator_loss", float(g_loss))
+                report("eval/generator_variance_loss", float(var_loss))
+                report("eval/generator_generator_mel_loss", float(mel_loss))
+                report("eval/generator_generator_adv_loss", float(adv_loss))
+                report("eval/generator_generator_feat_match_loss",
+                       float(feat_match_loss))
+                report("eval/generator_variance_dur_loss", float(dur_loss))
+                report("eval/generator_variance_pitch_loss", float(pitch_loss))
+                report("eval/generator_variance_energy_loss",
+                       float(energy_loss))
+
+                losses_dict["generator_loss"] = float(gen_loss)
+                losses_dict["generator_generator_loss"] = float(g_loss)
+                losses_dict["generator_variance_loss"] = float(var_loss)
+                losses_dict["generator_generator_mel_loss"] = float(mel_loss)
+                losses_dict["generator_generator_adv_loss"] = float(adv_loss)
+                losses_dict["generator_generator_feat_match_loss"] = float(
+                    feat_match_loss)
+                losses_dict["generator_variance_dur_loss"] = float(dur_loss)
+                losses_dict["generator_variance_pitch_loss"] = float(pitch_loss)
+                losses_dict["generator_variance_energy_loss"] = float(
+                    energy_loss)
+
+                if self.use_alignment_module == True:
+                    forwardsum_loss = self.criterion_forwardsum(
+                        log_p_attn, text_lengths, feats_lengths)
+                    align_loss = (
+                        forwardsum_loss + bin_loss) * self.lambda_align
+                    report("eval/generator_alignment_loss", float(align_loss))
+                    report("eval/generator_alignment_forwardsum_loss",
+                           float(forwardsum_loss))
+                    report("eval/generator_alignment_bin_loss", float(bin_loss))
+                    losses_dict["generator_alignment_loss"] = float(align_loss)
+                    losses_dict["generator_alignment_forwardsum_loss"] = float(
+                        forwardsum_loss)
+                    losses_dict["generator_alignment_bin_loss"] = float(
+                        bin_loss)
+
+                # reset cache
+                if self.model.reuse_cache_gen or not self.model.training:
+                    self.model._cache = None
+
+            # Disctiminator
+            elif turn == "discriminator":
+                # parse outputs
+                speech_hat_, _, _, start_idxs, *_ = outs
+                speech_ = get_segments(
+                    x=speech,
+                    start_idxs=start_idxs *
+                    self.model.generator.upsample_factor,
+                    segment_size=self.model.generator.segment_size *
+                    self.model.generator.upsample_factor, )
+
+                # calculate discriminator outputs
+                p_hat = self.model.discriminator(speech_hat_.detach())
+                p = self.model.discriminator(speech_)
+
+                # calculate losses
+                real_loss, fake_loss = self.criterion_dis_adv(p_hat, p)
+                dis_loss = real_loss + fake_loss
+
+                report("eval/real_loss", float(real_loss))
+                report("eval/fake_loss", float(fake_loss))
+                report("eval/discriminator_loss", float(dis_loss))
+                losses_dict["real_loss"] = float(real_loss)
+                losses_dict["fake_loss"] = float(fake_loss)
+                losses_dict["discriminator_loss"] = float(dis_loss)
+
+                # reset cache
+                if self.model.reuse_cache_dis or not self.model.training:
+                    self.model._cache = None
+
+        self.msg += ', '.join('{}: {:>.6f}'.format(k, v)
+                              for k, v in losses_dict.items())
+        self.logger.info(self.msg)

paddlespeech/t2s/models/jets/length_regulator.py

@@ -0,0 +1,67 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Generator module in JETS.
+
+This code is based on https://github.com/imdanboy/jets.
+
+"""
+import paddle
+import paddle.nn.functional as F
+from paddle import nn
+
+from paddlespeech.t2s.modules.masked_fill import masked_fill
+
+
+class GaussianUpsampling(nn.Layer):
+    """
+    Gaussian upsampling with fixed temperature as in:
+    https://arxiv.org/abs/2010.04301
+    """
+
+    def __init__(self, delta=0.1):
+        super().__init__()
+        self.delta = delta
+
+    def forward(self, hs, ds, h_masks=None, d_masks=None):
+        """
+        Args:
+            hs (Tensor): Batched hidden state to be expanded (B, T_text, adim)
+            ds (Tensor): Batched token duration (B, T_text)
+            h_masks (Tensor): Mask tensor (B,T_feats)
+            d_masks (Tensor): Mask tensor (B,T_text)
+        Returns:
+            Tensor: Expanded hidden state (B, T_feat, adim)
+        """
+        B = ds.shape[0]
+
+        if h_masks is None:
+            T_feats = paddle.to_tensor(ds.sum(), dtype="int32")
+        else:
+            T_feats = h_masks.shape[-1]
+        t = paddle.to_tensor(
+            paddle.arange(0, T_feats).unsqueeze(0).tile([B, 1]),
+            dtype="float32")
+        if h_masks is not None:
+            t = t * paddle.to_tensor(h_masks, dtype="float32")
+
+        c = ds.cumsum(axis=-1) - ds / 2
+        energy = -1 * self.delta * (t.unsqueeze(-1) - c.unsqueeze(1))**2
+        if d_masks is not None:
+            d_masks = ~(d_masks.unsqueeze(1))
+            d_masks.stop_gradient = True
+            d_masks = d_masks.tile([1, T_feats, 1])
+            energy = masked_fill(energy, d_masks, -float("inf"))
+        p_attn = F.softmax(energy, axis=2)  # (B, T_feats, T_text)
+        hs = paddle.matmul(p_attn, hs)
+        return hs

paddlespeech/t2s/modules/losses.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import math
+from typing import Tuple
 
 import librosa
 import numpy as np
@@ -19,8 +20,13 @@ import paddle
 from paddle import nn
 from paddle.nn import functional as F
 from scipy import signal
+from scipy.stats import betabinom
+from typeguard import check_argument_types
 
 from paddlespeech.t2s.modules.nets_utils import make_non_pad_mask
+from paddlespeech.t2s.modules.predictor.duration_predictor import (
+    DurationPredictorLoss,  # noqa: H301
+)
 
 
 # Losses for WaveRNN
@@ -1126,3 +1132,195 @@ class MLMLoss(nn.Layer):
                 text_masked_pos_reshape) / paddle.sum((text_masked_pos) + 1e-10)
 
         return mlm_loss, text_mlm_loss
+
+
+class VarianceLoss(nn.Layer):
+    def __init__(self, use_masking: bool=True,
+                 use_weighted_masking: bool=False):
+        """Initialize JETS variance loss module.
+        Args:
+            use_masking (bool): Whether to apply masking for padded part in loss
+                calculation.
+            use_weighted_masking (bool): Whether to weighted masking in loss
+                calculation.
+
+        """
+        assert check_argument_types()
+        super().__init__()
+
+        assert (use_masking != use_weighted_masking) or not use_masking
+        self.use_masking = use_masking
+        self.use_weighted_masking = use_weighted_masking
+
+        # define criterions
+        reduction = "none" if self.use_weighted_masking else "mean"
+        self.mse_criterion = nn.MSELoss(reduction=reduction)
+        self.duration_criterion = DurationPredictorLoss(reduction=reduction)
+
+    def forward(
+            self,
+            d_outs: paddle.Tensor,
+            ds: paddle.Tensor,
+            p_outs: paddle.Tensor,
+            ps: paddle.Tensor,
+            e_outs: paddle.Tensor,
+            es: paddle.Tensor,
+            ilens: paddle.Tensor,
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Calculate forward propagation.
+
+        Args:
+            d_outs (LongTensor): Batch of outputs of duration predictor (B, T_text).
+            ds (LongTensor): Batch of durations (B, T_text).
+            p_outs (Tensor): Batch of outputs of pitch predictor (B, T_text, 1).
+            ps (Tensor): Batch of target token-averaged pitch (B, T_text, 1).
+            e_outs (Tensor): Batch of outputs of energy predictor (B, T_text, 1).
+            es (Tensor): Batch of target token-averaged energy (B, T_text, 1).
+            ilens (LongTensor): Batch of the lengths of each input (B,).
+
+        Returns:
+            Tensor: Duration predictor loss value.
+            Tensor: Pitch predictor loss value.
+            Tensor: Energy predictor loss value.
+
+        """
+        # apply mask to remove padded part
+        if self.use_masking:
+            duration_masks = paddle.to_tensor(
+                make_non_pad_mask(ilens), place=ds.place)
+            d_outs = d_outs.masked_select(duration_masks)
+            ds = ds.masked_select(duration_masks)
+            pitch_masks = paddle.to_tensor(
+                make_non_pad_mask(ilens).unsqueeze(-1), place=ds.place)
+            p_outs = p_outs.masked_select(pitch_masks)
+            e_outs = e_outs.masked_select(pitch_masks)
+            ps = ps.masked_select(pitch_masks)
+            es = es.masked_select(pitch_masks)
+
+        # calculate loss
+        duration_loss = self.duration_criterion(d_outs, ds)
+        pitch_loss = self.mse_criterion(p_outs, ps)
+        energy_loss = self.mse_criterion(e_outs, es)
+
+        # make weighted mask and apply it
+        if self.use_weighted_masking:
+            duration_masks = paddle.to_tensor(
+                make_non_pad_mask(ilens), place=ds.place)
+            duration_weights = (duration_masks.float() /
+                                duration_masks.sum(dim=1, keepdim=True).float())
+            duration_weights /= ds.size(0)
+
+            # apply weight
+            duration_loss = (duration_loss.mul(duration_weights).masked_select(
+                duration_masks).sum())
+            pitch_masks = duration_masks.unsqueeze(-1)
+            pitch_weights = duration_weights.unsqueeze(-1)
+            pitch_loss = pitch_loss.mul(pitch_weights).masked_select(
+                pitch_masks).sum()
+            energy_loss = (
+                energy_loss.mul(pitch_weights).masked_select(pitch_masks).sum())
+
+        return duration_loss, pitch_loss, energy_loss
+
+
+class ForwardSumLoss(nn.Layer):
+    """
+    https://openreview.net/forum?id=0NQwnnwAORi
+    """
+
+    def __init__(self, cache_prior: bool=True):
+        """
+        Args:
+            cache_prior (bool): Whether to cache beta-binomial prior
+        """
+        super().__init__()
+        self.cache_prior = cache_prior
+        self._cache = {}
+
+    def forward(
+            self,
+            log_p_attn: paddle.Tensor,
+            ilens: paddle.Tensor,
+            olens: paddle.Tensor,
+            blank_prob: float=np.e**-1, ) -> paddle.Tensor:
+        """
+        Args:
+            log_p_attn (Tensor): Batch of log probability of attention matrix (B, T_feats, T_text).
+            ilens (Tensor): Batch of the lengths of each input (B,).
+            olens (Tensor): Batch of the lengths of each target (B,).
+            blank_prob (float): Blank symbol probability
+
+        Returns:
+            Tensor: forwardsum loss value.
+        """
+
+        B = log_p_attn.shape[0]
+        # add beta-binomial prior
+        bb_prior = self._generate_prior(ilens, olens)
+        bb_prior = paddle.to_tensor(
+            bb_prior, dtype=log_p_attn.dtype, place=log_p_attn.place)
+        log_p_attn = log_p_attn + bb_prior
+
+        # a row must be added to the attention matrix to account for blank token of CTC loss
+        # (B,T_feats,T_text+1)
+        log_p_attn_pd = F.pad(
+            log_p_attn, (0, 0, 0, 0, 1, 0), value=np.log(blank_prob))
+        loss = 0
+        for bidx in range(B):
+            # construct target sequnece.
+            # Every text token is mapped to a unique sequnece number.
+            target_seq = paddle.arange(
+                1, ilens[bidx] + 1, dtype="int32").unsqueeze(0)
+            cur_log_p_attn_pd = log_p_attn_pd[bidx, :olens[bidx], :ilens[
+                bidx] + 1].unsqueeze(1)  # (T_feats,1,T_text+1)
+            # The input of ctc_loss API need to be fixed
+            loss += F.ctc_loss(
+                log_probs=cur_log_p_attn_pd,
+                labels=target_seq,
+                input_lengths=olens[bidx:bidx + 1],
+                label_lengths=ilens[bidx:bidx + 1])
+        loss = loss / B
+
+        return loss
+
+    def _generate_prior(self, text_lengths, feats_lengths,
+                        w=1) -> paddle.Tensor:
+        """Generate alignment prior formulated as beta-binomial distribution
+
+        Args:
+            text_lengths (Tensor): Batch of the lengths of each input (B,).
+            feats_lengths (Tensor): Batch of the lengths of each target (B,).
+            w (float): Scaling factor; lower -> wider the width
+
+        Returns:
+            Tensor: Batched 2d static prior matrix (B, T_feats, T_text)   
+        """
+        B = len(text_lengths)
+        T_text = text_lengths.max()
+        T_feats = feats_lengths.max()
+
+        bb_prior = paddle.full((B, T_feats, T_text), fill_value=-np.inf)
+        for bidx in range(B):
+            T = feats_lengths[bidx].item()
+            N = text_lengths[bidx].item()
+
+            key = str(T) + ',' + str(N)
+            if self.cache_prior and key in self._cache:
+                prob = self._cache[key]
+            else:
+                alpha = w * np.arange(1, T + 1, dtype=float)  # (T,)
+                beta = w * np.array([T - t + 1 for t in alpha])
+                k = np.arange(N)
+                batched_k = k[..., None]  # (N,1)
+                prob = betabinom.pmf(batched_k, N, alpha, beta)  # (N,T)
+
+            # store cache
+            if self.cache_prior and key not in self._cache:
+                self._cache[key] = prob
+
+            prob = paddle.to_tensor(
+                prob, place=text_lengths.place, dtype="float32").transpose(
+                    (1, 0))  # -> (T,N)
+            bb_prior[bidx, :T, :N] = prob
+
+        return bb_prior