update tacotron2 voice cloning in aishell3 with new tacotron2, test=tts (#1419)

pull/1429/head
TianYuan 3 years ago committed by GitHub
parent a3ff7c0df1
commit 348a1a33bf
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

@ -1,94 +1,119 @@
# Tacotron2 + AISHELL-3 Voice Cloning
This example contains code used to train a [Tacotron2 ](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in Tacotron2 because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
2. Synthesizer: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of Tacotron2 which will be concated with encoder outputs.
3. Vocoder: We use WaveFlow as the neural Vocoder, refer to [waveflow](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/ljspeech/voc0).
This example contains code used to train a [Tacotron2](https://arxiv.org/abs/1712.05884) model with [AISHELL-3](http://www.aishelltech.com/aishell_3). The trained model can be used in Voice Cloning Task, We refer to the model structure of [Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis](https://arxiv.org/pdf/1806.04558.pdf). The general steps are as follows:
1. Speaker Encoder: We use Speaker Verification to train a speaker encoder. Datasets used in this task are different from those used in `Tacotron2` because the transcriptions are not needed, we use more datasets, refer to [ge2e](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/ge2e).
2. Synthesizer: We use the trained speaker encoder to generate speaker embedding for each sentence in AISHELL-3. This embedding is an extra input of `Tacotron2` which will be concated with encoder outputs.
3. Vocoder: We use [Parallel Wave GAN](http://arxiv.org/abs/1910.11480) as the neural Vocoder, refer to [voc1](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/voc1).
## Dataset
### Download and Extract
Download AISHELL-3.
```bash
wget https://www.openslr.org/resources/93/data_aishell3.tgz
```
Extract AISHELL-3.
```bash
mkdir data_aishell3
tar zxvf data_aishell3.tgz -C data_aishell3
```
### Get MFA Result and Extract
We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get durations for aishell3_fastspeech2.
You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
## Pretrained GE2E Model
We use pretrained GE2E model to generate speaker embedding for each sentence.
Download pretrained GE2E model from here [ge2e_ckpt_0.3.zip](https://bj.bcebos.com/paddlespeech/Parakeet/released_models/ge2e/ge2e_ckpt_0.3.zip), and `unzip` it.
## Get Started
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the MFA result of AISHELL-3 is `./alignment`.
Assume the path to the pretrained ge2e model is `ge2e_ckpt_path=./ge2e_ckpt_0.3/step-3000000`
Assume the path to the MFA result of AISHELL-3 is `./aishell3_alignment_tone`.
Assume the path to the pretrained ge2e model is `./ge2e_ckpt_0.3`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. start a voice cloning inference.
4. synthesize waveform from `metadata.jsonl`.
5. start a voice cloning inference.
```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, run the following command will only preprocess the dataset.
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
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${input} ${preprocess_path} ${alignment} ${ge2e_ckpt_path}
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${conf_path} ${ge2e_ckpt_path}
```
#### Generate Speaker Embedding
Use pretrained GE2E (speaker encoder) to generate speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
```bash
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
python3 ${BIN_DIR}/../ge2e/inference.py \
--input=${input} \
--output=${preprocess_path}/embed \
--ngpu=1 \
--checkpoint_path=${ge2e_ckpt_path}
fi
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
├── embed
│ ├── SSB0005
│ ├── SSB0009
│ ├── ...
│ └── ...
├── phone_id_map.txt
├── speaker_id_map.txt
├── test
│ ├── norm
│ └── raw
└── train
├── norm
├── raw
└── speech_stats.npy
```
The `embed` contains the generated speaker embedding for each sentence in AISHELL-3, which has the same file structure with wav files and the format is `.npy`.
The computing time of utterance embedding can be x hours.
#### Process Wav
There is silence in the edge of AISHELL-3's wavs, and the audio amplitude is very small, so, we need to remove the silence and normalize the audio. You can the silence remove method based on volume or energy, but the effect is not very good, We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get the alignment of text and speech, then utilize the alignment results to remove the silence.
We use Montreal Force Aligner 1.0. The label in aishell3 includes pinyinso the lexicon we provided to MFA is pinyin rather than Chinese characters. And the prosody marks(`$` and `%`) need to be removed. You should preprocess the dataset into the format which MFA needs, the texts have the same name with wavs and have the suffix `.lab`.
We use [lexicon.txt](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/paddlespeech/t2s/exps/voice_cloning/tacotron2_ge2e/lexicon.txt) as the lexicon.
The dataset is split into 3 parts, namely `train`, `dev`, and` test`, each of which contains a `norm` and `raw` subfolder. The raw folder contains speech features of each utterance, while the norm folder contains normalized ones. The statistics used to normalize features are computed from the training set, which is located in `dump/train/*_stats.npy`.
You can download the alignment results from here [alignment_aishell3.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/alignment_aishell3.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, speech_lengths, durations, the path of speech features, speaker, and id of each utterance.
The preprocessing step is very similar to that one of [tts0](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/tts0), but there is one more `ge2e/inference` step here.
### Model Training
`./local/train.sh` calls `${BIN_DIR}/train.py`.
```bash
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "Process wav ..."
python3 ${BIN_DIR}/process_wav.py \
--input=${input}/wav \
--output=${preprocess_path}/normalized_wav \
--alignment=${alignment}
fi
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
The training step is very similar to that one of [tts0](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/tts0), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/train.py`.
#### Preprocess Transcription
We revert the transcription into `phones` and `tones`. It is worth noting that our processing here is different from that used for MFA, we separated the tones. This is a processing method, of course, you can only segment initials and vowels.
### Synthesizing
We use [parallel wavegan](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/aishell3/voc1) as the neural vocoder.
Download pretrained parallel wavegan model from [pwg_aishell3_ckpt_0.5.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwg_aishell3_ckpt_0.5.zip) and unzip it.
```bash
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
python3 ${BIN_DIR}/preprocess_transcription.py \
--input=${input} \
--output=${preprocess_path}
fi
unzip pwg_aishell3_ckpt_0.5.zip
```
The default input is `~/datasets/data_aishell3/train`which contains `label_train-set.txt`, the processed results are `metadata.yaml` and `metadata.pickle`. the former is a text format for easy viewing, and the latter is a binary format for direct reading.
#### Extract Mel
```python
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
python3 ${BIN_DIR}/extract_mel.py \
--input=${preprocess_path}/normalized_wav \
--output=${preprocess_path}/mel
fi
Parallel WaveGAN checkpoint contains files listed below.
```text
pwg_aishell3_ckpt_0.5
├── default.yaml # default config used to train parallel wavegan
├── feats_stats.npy # statistics used to normalize spectrogram when training parallel wavegan
└── snapshot_iter_1000000.pdz # generator parameters of parallel wavegan
```
### Model Training
`./local/synthesize.sh` calls `${BIN_DIR}/../synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${preprocess_path} ${train_output_path}
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
The synthesizing step is very similar to that one of [tts0](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/csmsc/tts0), but we should set `--voice-cloning=True` when calling `${BIN_DIR}/../synthesize.py`.
Our model removes stop token prediction in Tacotron2, because of the problem of the extremely unbalanced proportion of positive and negative samples of stop token prediction, and it's very sensitive to the clip of audio silence. We use the last symbol from the highest point of attention to the encoder side as the termination condition.
In addition, to accelerate the convergence of the model, we add `guided attention loss` to induce the alignment between encoder and decoder to show diagonal lines faster.
### Voice Cloning
Assume there are some reference audios in `./ref_audio`
```text
ref_audio
├── 001238.wav
├── LJ015-0254.wav
└── audio_self_test.mp3
```
`./local/voice_cloning.sh` calls `${BIN_DIR}/../voice_cloning.py`
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${ge2e_params_path} ${tacotron2_params_path} ${waveflow_params_path} ${vc_input} ${vc_output}
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${ref_audio_dir}
```
## Pretrained Model
[tacotron2_aishell3_ckpt_0.3.zip](https://paddlespeech.bj.bcebos.com/Parakeet/released_models/tacotron2/tacotron2_aishell3_ckpt_0.3.zip).

@ -0,0 +1,86 @@
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # sr
n_fft: 2048 # FFT size (samples).
n_shift: 300 # Hop size (samples). 12.5ms
win_length: 1200 # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
# Only used for feats_type != raw
fmin: 80 # Minimum frequency of Mel basis.
fmax: 7600 # Maximum frequency of Mel basis.
n_mels: 80 # The number of mel basis.
###########################################################
# DATA SETTING #
###########################################################
batch_size: 64
num_workers: 2
###########################################################
# MODEL SETTING #
###########################################################
model: # keyword arguments for the selected model
embed_dim: 512 # char or phn embedding dimension
elayers: 1 # number of blstm layers in encoder
eunits: 512 # number of blstm units
econv_layers: 3 # number of convolutional layers in encoder
econv_chans: 512 # number of channels in convolutional layer
econv_filts: 5 # filter size of convolutional layer
atype: location # attention function type
adim: 512 # attention dimension
aconv_chans: 32 # number of channels in convolutional layer of attention
aconv_filts: 15 # filter size of convolutional layer of attention
cumulate_att_w: True # whether to cumulate attention weight
dlayers: 2 # number of lstm layers in decoder
dunits: 1024 # number of lstm units in decoder
prenet_layers: 2 # number of layers in prenet
prenet_units: 256 # number of units in prenet
postnet_layers: 5 # number of layers in postnet
postnet_chans: 512 # number of channels in postnet
postnet_filts: 5 # filter size of postnet layer
output_activation: null # activation function for the final output
use_batch_norm: True # whether to use batch normalization in encoder
use_concate: True # whether to concatenate encoder embedding with decoder outputs
use_residual: False # whether to use residual connection in encoder
dropout_rate: 0.5 # dropout rate
zoneout_rate: 0.1 # zoneout rate
reduction_factor: 1 # reduction factor
spk_embed_dim: 256 # speaker embedding dimension
spk_embed_integration_type: concat # how to integrate speaker embedding
###########################################################
# UPDATER SETTING #
###########################################################
updater:
use_masking: True # whether to apply masking for padded part in loss calculation
bce_pos_weight: 5.0 # weight of positive sample in binary cross entropy calculation
use_guided_attn_loss: True # whether to use guided attention loss
guided_attn_loss_sigma: 0.4 # sigma of guided attention loss
guided_attn_loss_lambda: 1.0 # strength of guided attention loss
##########################################################
# OPTIMIZER SETTING #
##########################################################
optimizer:
optim: adam # optimizer type
learning_rate: 1.0e-03 # learning rate
epsilon: 1.0e-06 # epsilon
weight_decay: 0.0 # weight decay coefficient
###########################################################
# TRAINING SETTING #
###########################################################
max_epoch: 200
num_snapshots: 5
###########################################################
# OTHER SETTING #
###########################################################
seed: 42

@ -1,36 +1,72 @@
#!/bin/bash
stage=0
stage=3
stop_stage=100
input=$1
preprocess_path=$2
alignment=$3
ge2e_ckpt_path=$4
config_path=$1
ge2e_ckpt_path=$2
# gen speaker embedding
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
python3 ${MAIN_ROOT}/paddlespeech/vector/exps/ge2e/inference.py \
--input=${input}/wav \
--output=${preprocess_path}/embed \
--input=~/datasets/data_aishell3/train/wav/ \
--output=dump/embed \
--checkpoint_path=${ge2e_ckpt_path}
fi
# copy from tts3/preprocess
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "Process wav ..."
python3 ${BIN_DIR}/process_wav.py \
--input=${input}/wav \
--output=${preprocess_path}/normalized_wav \
--alignment=${alignment}
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./aishell3_alignment_tone \
--output durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
python3 ${BIN_DIR}/preprocess_transcription.py \
--input=${input} \
--output=${preprocess_path}
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/preprocess.py \
--dataset=aishell3 \
--rootdir=~/datasets/data_aishell3/ \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--num-cpu=20 \
--cut-sil=True \
--spk_emb_dir=dump/embed
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
python3 ${BIN_DIR}/extract_mel.py \
--input=${preprocess_path}/normalized_wav \
--output=${preprocess_path}/mel
# 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="speech"
fi
if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
# normalize and covert phone 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 \
--speech-stats=dump/train/speech_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 \
--speech-stats=dump/train/speech_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 \
--speech-stats=dump/train/speech_stats.npy \
--phones-dict=dump/phone_id_map.txt \
--speaker-dict=dump/speaker_id_map.txt
fi

@ -0,0 +1,22 @@
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/../synthesize.py \
--am=tacotron2_aishell3 \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--voc=pwgan_aishell3 \
--voc_config=pwg_aishell3_ckpt_0.5/default.yaml \
--voc_ckpt=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
--voc_stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
--test_metadata=dump/test/norm/metadata.jsonl \
--output_dir=${train_output_path}/test \
--phones_dict=dump/phone_id_map.txt \
--speaker_dict=dump/speaker_id_map.txt \
--voice-cloning=True

@ -1,9 +1,13 @@
#!/bin/bash
preprocess_path=$1
config_path=$1
train_output_path=$2
python3 ${BIN_DIR}/train.py \
--data=${preprocess_path} \
--output=${train_output_path} \
--ngpu=1
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=2 \
--phones-dict=dump/phone_id_map.txt \
--voice-cloning=True

@ -1,14 +1,24 @@
#!/bin/bash
ge2e_params_path=$1
tacotron2_params_path=$2
waveflow_params_path=$3
vc_input=$4
vc_output=$5
config_path=$1
train_output_path=$2
ckpt_name=$3
ge2e_params_path=$4
ref_audio_dir=$5
python3 ${BIN_DIR}/voice_cloning.py \
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/../voice_cloning.py \
--am=tacotron2_aishell3 \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--voc=pwgan_aishell3 \
--voc_config=pwg_aishell3_ckpt_0.5/default.yaml \
--voc_ckpt=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
--voc_stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
--ge2e_params_path=${ge2e_params_path} \
--tacotron2_params_path=${tacotron2_params_path} \
--waveflow_params_path=${waveflow_params_path} \
--input-dir=${vc_input} \
--output-dir=${vc_output}
--text="凯莫瑞安联合体的经济崩溃迫在眉睫。" \
--input-dir=${ref_audio_dir} \
--output-dir=${train_output_path}/vc_syn \
--phones-dict=dump/phone_id_map.txt

@ -9,5 +9,5 @@ export PYTHONDONTWRITEBYTECODE=1
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=voice_cloning/tacotron2_ge2e
MODEL=new_tacotron2
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/${MODEL}

@ -3,25 +3,20 @@
set -e
source path.sh
gpus=0
gpus=0,1
stage=0
stop_stage=100
input=~/datasets/data_aishell3/train
preprocess_path=dump
alignment=./alignment
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_482.pdz
ref_audio_dir=ref_audio
# not include ".pdparams" here
ge2e_ckpt_path=./ge2e_ckpt_0.3/step-3000000
train_output_path=output
# include ".pdparams" here
ge2e_params_path=${ge2e_ckpt_path}.pdparams
tacotron2_params_path=${train_output_path}/checkpoints/step-1000.pdparams
# pretrained model
# tacotron2_params_path=./tacotron2_aishell3_ckpt_0.3/step-450000.pdparams
waveflow_params_path=./waveflow_ljspeech_ckpt_0.3/step-2000000.pdparams
vc_input=ref_audio
vc_output=syn_audio
# with the following command, you can choose the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
@ -30,15 +25,20 @@ source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${input} ${preprocess_path} ${alignment} ${ge2e_ckpt_path} || exit -1
CUDA_VISIBLE_DEVICES=${gpus} ./local/preprocess.sh ${conf_path} ${ge2e_ckpt_path} || exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${preprocess_path} ${train_output_path} || exit -1
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${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/voice_cloning.sh ${ge2e_params_path} ${tacotron2_params_path} ${waveflow_params_path} ${vc_input} ${vc_output} || exit -1
# synthesize, vocoder is pwgan
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# synthesize, vocoder is pwgan
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${ref_audio_dir} || exit -1
fi

@ -114,7 +114,7 @@ ref_audio
├── LJ015-0254.wav
└── audio_self_test.mp3
```
`./local/voice_cloning.sh` calls `${BIN_DIR}/voice_cloning.py`
`./local/voice_cloning.sh` calls `${BIN_DIR}/../voice_cloning.py`
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/voice_cloning.sh ${conf_path} ${train_output_path} ${ckpt_name} ${ge2e_params_path} ${ref_audio_dir}

@ -8,13 +8,15 @@ ref_audio_dir=$5
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/voice_cloning.py \
--fastspeech2-config=${config_path} \
--fastspeech2-checkpoint=${train_output_path}/checkpoints/${ckpt_name} \
--fastspeech2-stat=dump/train/speech_stats.npy \
--pwg-config=pwg_aishell3_ckpt_0.5/default.yaml \
--pwg-checkpoint=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
--pwg-stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
python3 ${BIN_DIR}/../voice_cloning.py \
--am=fastspeech2_aishell3 \
--am_config=${config_path} \
--am_ckpt=${train_output_path}/checkpoints/${ckpt_name} \
--am_stat=dump/train/speech_stats.npy \
--voc=pwgan_aishell3 \
--voc_config=pwg_aishell3_ckpt_0.5/default.yaml \
--voc_ckpt=pwg_aishell3_ckpt_0.5/snapshot_iter_1000000.pdz \
--voc_stat=pwg_aishell3_ckpt_0.5/feats_stats.npy \
--ge2e_params_path=${ge2e_params_path} \
--text="凯莫瑞安联合体的经济崩溃迫在眉睫。" \
--input-dir=${ref_audio_dir} \

@ -44,15 +44,13 @@ dump
│ ├── norm
│ └── raw
└── train
├── energy_stats.npy
├── norm
├── pitch_stats.npy
├── raw
└── speech_stats.npy
```
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 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/*_stats.npy`.
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 speech 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/*_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, speech_lengths, durations, the path of speech features, the path of pitch features, the path of energy features, speaker, and the id of each utterance.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains phones, text_lengths, speech_lengths, durations, the path of speech features, speaker, and the id of each utterance.
### Model Training
```bash

@ -46,6 +46,47 @@ def tacotron2_single_spk_batch_fn(examples):
return batch
def tacotron2_multi_spk_batch_fn(examples):
# fields = ["text", "text_lengths", "speech", "speech_lengths"]
text = [np.array(item["text"], dtype=np.int64) for item in examples]
speech = [np.array(item["speech"], dtype=np.float32) for item in examples]
text_lengths = [
np.array(item["text_lengths"], dtype=np.int64) for item in examples
]
speech_lengths = [
np.array(item["speech_lengths"], dtype=np.int64) for item in examples
]
text = batch_sequences(text)
speech = batch_sequences(speech)
# convert each batch to paddle.Tensor
text = paddle.to_tensor(text)
speech = paddle.to_tensor(speech)
text_lengths = paddle.to_tensor(text_lengths)
speech_lengths = paddle.to_tensor(speech_lengths)
batch = {
"text": text,
"text_lengths": text_lengths,
"speech": speech,
"speech_lengths": speech_lengths,
}
# 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
def speedyspeech_single_spk_batch_fn(examples):
# fields = ["phones", "tones", "num_phones", "num_frames", "feats", "durations"]
phones = [np.array(item["phones"], dtype=np.int64) for item in examples]

@ -29,6 +29,7 @@ from paddlespeech.t2s.datasets.preprocess_utils import merge_silence
from paddlespeech.t2s.models.fastspeech2 import FastSpeech2
from paddlespeech.t2s.models.fastspeech2 import StyleFastSpeech2Inference
from paddlespeech.t2s.modules.normalizer import ZScore
from paddlespeech.t2s.utils import str2bool
def evaluate(args, fastspeech2_config):
@ -196,9 +197,6 @@ def main():
parser.add_argument(
"--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--cut-sil",
type=str2bool,

@ -35,6 +35,7 @@ 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],
@ -203,9 +204,6 @@ def main():
parser.add_argument(
"--num-cpu", type=int, default=1, help="number of process.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--cut-sil",
type=str2bool,

@ -38,6 +38,7 @@ from paddlespeech.t2s.training.extensions.visualizer import VisualDL
from paddlespeech.t2s.training.optimizer import build_optimizers
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):
@ -182,9 +183,6 @@ def main():
default=None,
help="speaker id map file for multiple speaker model.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--voice-cloning",
type=str2bool,

@ -41,6 +41,7 @@ from paddlespeech.t2s.training.extensions.snapshot import Snapshot
from paddlespeech.t2s.training.extensions.visualizer import VisualDL
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):
@ -204,8 +205,6 @@ def train_sp(args, config):
def main():
# parse args and config and redirect to train_sp
def str2bool(str):
return True if str.lower() == 'true' else False
parser = argparse.ArgumentParser(
description="Train a ParallelWaveGAN model.")

@ -30,6 +30,7 @@ from yacs.config import CfgNode
from paddlespeech.t2s.data.get_feats import LogMelFBank
from paddlespeech.t2s.datasets.preprocess_utils import get_phn_dur
from paddlespeech.t2s.datasets.preprocess_utils import merge_silence
from paddlespeech.t2s.utils import str2bool
def process_sentence(config: Dict[str, Any],
@ -165,9 +166,6 @@ def main():
parser.add_argument(
"--dur-file", default=None, type=str, help="path to durations.txt.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--cut-sil",
type=str2bool,

@ -33,6 +33,7 @@ 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],
@ -179,9 +180,6 @@ def main():
parser.add_argument(
"--num-cpu", type=int, default=1, help="number of process.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--cut-sil",
type=str2bool,

@ -27,6 +27,7 @@ from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from yacs.config import CfgNode
from paddlespeech.t2s.datasets.am_batch_fn import tacotron2_multi_spk_batch_fn
from paddlespeech.t2s.datasets.am_batch_fn import tacotron2_single_spk_batch_fn
from paddlespeech.t2s.datasets.data_table import DataTable
from paddlespeech.t2s.models.new_tacotron2 import Tacotron2
@ -37,6 +38,7 @@ from paddlespeech.t2s.training.extensions.visualizer import VisualDL
from paddlespeech.t2s.training.optimizer import build_optimizers
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):
@ -60,33 +62,38 @@ def train_sp(args, config):
# dataloader has been too verbose
logging.getLogger("DataLoader").disabled = True
fields = [
"text",
"text_lengths",
"speech",
"speech_lengths",
]
converters = {
"speech": np.load,
}
if args.voice_cloning:
print("Training voice cloning!")
collate_fn = tacotron2_multi_spk_batch_fn
fields += ["spk_emb"]
converters["spk_emb"] = np.load
else:
print("single speaker tacotron2!")
collate_fn = tacotron2_single_spk_batch_fn
# 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=[
"text",
"text_lengths",
"speech",
"speech_lengths",
],
converters={
"speech": np.load,
}, )
fields=fields,
converters=converters, )
with jsonlines.open(args.dev_metadata, 'r') as reader:
dev_metadata = list(reader)
dev_dataset = DataTable(
data=dev_metadata,
fields=[
"text",
"text_lengths",
"speech",
"speech_lengths",
],
converters={
"speech": np.load,
}, )
fields=fields,
converters=converters, )
# collate function and dataloader
train_sampler = DistributedBatchSampler(
@ -100,7 +107,7 @@ def train_sp(args, config):
train_dataloader = DataLoader(
train_dataset,
batch_sampler=train_sampler,
collate_fn=tacotron2_single_spk_batch_fn,
collate_fn=collate_fn,
num_workers=config.num_workers)
dev_dataloader = DataLoader(
@ -108,7 +115,7 @@ def train_sp(args, config):
shuffle=False,
drop_last=False,
batch_size=config.batch_size,
collate_fn=tacotron2_single_spk_batch_fn,
collate_fn=collate_fn,
num_workers=config.num_workers)
print("dataloaders done!")
@ -166,6 +173,12 @@ def main():
parser.add_argument(
"--phones-dict", type=str, default=None, help="phone vocabulary file.")
parser.add_argument(
"--voice-cloning",
type=str2bool,
default=False,
help="whether training voice cloning model.")
args = parser.parse_args()
with open(args.config) as f:

@ -30,6 +30,7 @@ from paddlespeech.t2s.frontend.zh_frontend import Frontend
from paddlespeech.t2s.models.speedyspeech import SpeedySpeech
from paddlespeech.t2s.models.speedyspeech import SpeedySpeechInference
from paddlespeech.t2s.modules.normalizer import ZScore
from paddlespeech.t2s.utils import str2bool
def evaluate(args, speedyspeech_config):
@ -213,9 +214,6 @@ def main():
parser.add_argument(
"--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--cut-sil",
type=str2bool,

@ -23,6 +23,7 @@ from sklearn.preprocessing import StandardScaler
from tqdm import tqdm
from paddlespeech.t2s.datasets.data_table import DataTable
from paddlespeech.t2s.utils import str2bool
def main():
@ -55,9 +56,6 @@ def main():
default=1,
help="logging level. higher is more logging. (default=1)")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--use-relative-path",
type=str2bool,

@ -33,6 +33,7 @@ from paddlespeech.t2s.datasets.preprocess_utils import get_phn_dur
from paddlespeech.t2s.datasets.preprocess_utils import get_phones_tones
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],
@ -190,9 +191,6 @@ def main():
parser.add_argument(
"--num-cpu", type=int, default=1, help="number of process.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--cut-sil",
type=str2bool,

@ -38,6 +38,7 @@ from paddlespeech.t2s.training.extensions.visualizer import VisualDL
from paddlespeech.t2s.training.optimizer import build_optimizers
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):
@ -186,9 +187,6 @@ def main():
parser.add_argument(
"--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--use-relative-path",
type=str2bool,

@ -25,6 +25,7 @@ from yacs.config import CfgNode
from paddlespeech.s2t.utils.dynamic_import import dynamic_import
from paddlespeech.t2s.datasets.data_table import DataTable
from paddlespeech.t2s.modules.normalizer import ZScore
from paddlespeech.t2s.utils import str2bool
model_alias = {
# acoustic model
@ -97,6 +98,9 @@ def evaluate(args):
fields = ["utt_id", "phones", "tones"]
elif am_name == 'tacotron2':
fields = ["utt_id", "text"]
if args.voice_cloning:
print("voice cloning!")
fields += ["spk_emb"]
test_dataset = DataTable(data=test_metadata, fields=fields)
@ -178,7 +182,11 @@ def evaluate(args):
mel = am_inference(phone_ids, tone_ids)
elif am_name == 'tacotron2':
phone_ids = paddle.to_tensor(datum["text"])
mel = am_inference(phone_ids)
spk_emb = None
# multi speaker
if args.voice_cloning and "spk_emb" in datum:
spk_emb = paddle.to_tensor(np.load(datum["spk_emb"]))
mel = am_inference(phone_ids, spk_emb=spk_emb)
# vocoder
wav = voc_inference(mel)
sf.write(
@ -199,7 +207,7 @@ def main():
default='fastspeech2_csmsc',
choices=[
'speedyspeech_csmsc', 'fastspeech2_csmsc', 'fastspeech2_ljspeech',
'fastspeech2_aishell3', 'fastspeech2_vctk', 'tacotron2_csmsc'
'fastspeech2_aishell3', 'fastspeech2_vctk', 'tacotron2_csmsc', 'tacotron2_aishell3'
],
help='Choose acoustic model type of tts task.')
parser.add_argument(
@ -225,9 +233,6 @@ def main():
parser.add_argument(
"--speaker_dict", type=str, default=None, help="speaker id map file.")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--voice-cloning",
type=str2bool,

@ -1,13 +0,0 @@
# Copyright (c) 2020 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.

@ -1,75 +0,0 @@
# Copyright (c) 2020 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 yacs.config import CfgNode as CN
_C = CN()
_C.data = CN(
dict(
batch_size=32, # batch size
valid_size=64, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=1024, # fft frame size
win_length=1024, # window size
hop_length=256, # hop size between ajacent frame
fmax=8000, # Hz, max frequency when converting to mel
fmin=0, # Hz, min frequency when converting to mel
n_mels=80, # mel bands
padding_idx=0, # text embedding's padding index
))
_C.model = CN(
dict(
vocab_size=37, # set this according to the frontend's vocab size
n_tones=None,
reduction_factor=1, # reduction factor
d_encoder=512, # embedding & encoder's internal size
encoder_conv_layers=3, # number of conv layer in tacotron2 encoder
encoder_kernel_size=5, # kernel size of conv layers in tacotron2 encoder
d_prenet=256, # hidden size of decoder prenet
d_attention_rnn=1024, # hidden size of the first rnn layer in tacotron2 decoder
d_decoder_rnn=1024, # hidden size of the second rnn layer in tacotron2 decoder
d_attention=128, # hidden size of decoder location linear layer
attention_filters=32, # number of filter in decoder location conv layer
attention_kernel_size=31, # kernel size of decoder location conv layer
d_postnet=512, # hidden size of decoder postnet
postnet_kernel_size=5, # kernel size of conv layers in postnet
postnet_conv_layers=5, # number of conv layer in decoder postnet
p_encoder_dropout=0.5, # droput probability in encoder
p_prenet_dropout=0.5, # droput probability in decoder prenet
p_attention_dropout=0.1, # droput probability of first rnn layer in decoder
p_decoder_dropout=0.1, # droput probability of second rnn layer in decoder
p_postnet_dropout=0.5, # droput probability in decoder postnet
d_global_condition=None,
use_stop_token=True, # wherther to use binary classifier to predict when to stop
use_guided_attention_loss=False, # whether to use guided attention loss
guided_attention_loss_sigma=0.2 # sigma in guided attention loss
))
_C.training = CN(
dict(
lr=1e-3, # learning rate
weight_decay=1e-6, # the coeff of weight decay
grad_clip_thresh=1.0, # the clip norm of grad clip.
plot_interval=1000, # plot attention and spectrogram
valid_interval=1000, # validation
save_interval=1000, # checkpoint
max_iteration=500000, # max iteration to train
))
def get_cfg_defaults():
"""Get a yacs CfgNode object with default values for my_project."""
# Return a clone so that the defaults will not be altered
# This is for the "local variable" use pattern
return _C.clone()

@ -1,91 +0,0 @@
# Copyright (c) 2020 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 pickle
from pathlib import Path
import numpy as np
from paddle.io import Dataset
from paddlespeech.t2s.data.batch import batch_spec
from paddlespeech.t2s.data.batch import batch_text_id
class LJSpeech(Dataset):
"""A simple dataset adaptor for the processed ljspeech dataset."""
def __init__(self, root):
self.root = Path(root).expanduser()
records = []
with open(self.root / "metadata.pkl", 'rb') as f:
metadata = pickle.load(f)
for mel_name, text, ids in metadata:
mel_name = self.root / "mel" / (mel_name + ".npy")
records.append((mel_name, text, ids))
self.records = records
def __getitem__(self, i):
mel_name, _, ids = self.records[i]
mel = np.load(mel_name)
return ids, mel
def __len__(self):
return len(self.records)
class LJSpeechCollector(object):
"""A simple callable to batch LJSpeech examples."""
def __init__(self, padding_idx=0, padding_value=0., padding_stop_token=1.0):
self.padding_idx = padding_idx
self.padding_value = padding_value
self.padding_stop_token = padding_stop_token
def __call__(self, examples):
texts = []
mels = []
text_lens = []
mel_lens = []
for data in examples:
text, mel = data
text = np.array(text, dtype=np.int64)
text_lens.append(len(text))
mels.append(mel)
texts.append(text)
mel_lens.append(mel.shape[1])
# Sort by text_len in descending order
texts = [
i for i, _ in sorted(
zip(texts, text_lens), key=lambda x: x[1], reverse=True)
]
mels = [
i for i, _ in sorted(
zip(mels, text_lens), key=lambda x: x[1], reverse=True)
]
mel_lens = [
i for i, _ in sorted(
zip(mel_lens, text_lens), key=lambda x: x[1], reverse=True)
]
mel_lens = np.array(mel_lens, dtype=np.int64)
text_lens = np.array(sorted(text_lens, reverse=True), dtype=np.int64)
# Pad sequence with largest len of the batch
texts, _ = batch_text_id(texts, pad_id=self.padding_idx)
mels, _ = batch_spec(mels, pad_value=self.padding_value)
mels = np.transpose(mels, axes=(0, 2, 1))
return texts, mels, text_lens, mel_lens

@ -1,98 +0,0 @@
# Copyright (c) 2020 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
import pickle
from pathlib import Path
import numpy as np
import tqdm
from paddlespeech.t2s.audio import AudioProcessor
from paddlespeech.t2s.audio import LogMagnitude
from paddlespeech.t2s.datasets import LJSpeechMetaData
from paddlespeech.t2s.exps.tacotron2.config import get_cfg_defaults
from paddlespeech.t2s.frontend import EnglishCharacter
def create_dataset(config, source_path, target_path, verbose=False):
# create output dir
target_path = Path(target_path).expanduser()
mel_path = target_path / "mel"
os.makedirs(mel_path, exist_ok=True)
meta_data = LJSpeechMetaData(source_path)
frontend = EnglishCharacter()
processor = AudioProcessor(
sample_rate=config.data.sample_rate,
n_fft=config.data.n_fft,
n_mels=config.data.n_mels,
win_length=config.data.win_length,
hop_length=config.data.hop_length,
fmax=config.data.fmax,
fmin=config.data.fmin)
normalizer = LogMagnitude()
records = []
for (fname, text, _) in tqdm.tqdm(meta_data):
wav = processor.read_wav(fname)
mel = processor.mel_spectrogram(wav)
mel = normalizer.transform(mel)
ids = frontend(text)
mel_name = os.path.splitext(os.path.basename(fname))[0]
# save mel spectrogram
records.append((mel_name, text, ids))
np.save(mel_path / mel_name, mel)
if verbose:
print("save mel spectrograms into {}".format(mel_path))
# save meta data as pickle archive
with open(target_path / "metadata.pkl", 'wb') as f:
pickle.dump(records, f)
if verbose:
print("saved metadata into {}".format(target_path / "metadata.pkl"))
print("Done.")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="create dataset")
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--input", type=str, help="path of the ljspeech dataset")
parser.add_argument(
"--output", type=str, help="path to save output dataset")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
config = get_cfg_defaults()
args = parser.parse_args()
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)
config.freeze()
print(config.data)
create_dataset(config, args.input, args.output, args.verbose)

File diff suppressed because one or more lines are too long

@ -1,103 +0,0 @@
# Copyright (c) 2020 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 numpy as np
import paddle
from matplotlib import pyplot as plt
from paddlespeech.t2s.exps.tacotron2.config import get_cfg_defaults
from paddlespeech.t2s.frontend import EnglishCharacter
from paddlespeech.t2s.models.tacotron2 import Tacotron2
from paddlespeech.t2s.utils import display
def main(config, args):
if args.ngpu == 0:
paddle.set_device("cpu")
elif args.ngpu > 0:
paddle.set_device("gpu")
else:
print("ngpu should >= 0 !")
# model
frontend = EnglishCharacter()
model = Tacotron2.from_pretrained(config, args.checkpoint_path)
model.eval()
# inputs
input_path = Path(args.input).expanduser()
sentences = []
with open(input_path, "rt") as f:
for line in f:
line_list = line.strip().split()
utt_id = line_list[0]
sentence = " ".join(line_list[1:])
sentences.append((utt_id, sentence))
if args.output is None:
output_dir = input_path.parent / "synthesis"
else:
output_dir = Path(args.output).expanduser()
output_dir.mkdir(exist_ok=True)
for i, sentence in enumerate(sentences):
sentence = paddle.to_tensor(frontend(sentence)).unsqueeze(0)
outputs = model.infer(sentence)
mel_output = outputs["mel_outputs_postnet"][0].numpy().T
alignment = outputs["alignments"][0].numpy().T
np.save(str(output_dir / f"sentence_{i}"), mel_output)
display.plot_alignment(alignment)
plt.savefig(str(output_dir / f"sentence_{i}.png"))
if args.verbose:
print("spectrogram saved at {}".format(output_dir /
f"sentence_{i}.npy"))
if __name__ == "__main__":
config = get_cfg_defaults()
parser = argparse.ArgumentParser(
description="generate mel spectrogram with TransformerTTS.")
parser.add_argument(
"--config",
type=str,
metavar="FILE",
help="extra config to overwrite the default config")
parser.add_argument(
"--checkpoint_path", type=str, help="path of the checkpoint to load.")
parser.add_argument("--input", type=str, help="path of the text sentences")
parser.add_argument("--output", type=str, help="path to save outputs")
parser.add_argument(
"--ngpu", type=int, default=1, help="if ngpu == 0, use cpu.")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
parser.add_argument(
"-v", "--verbose", action="store_true", help="print msg")
args = parser.parse_args()
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)
config.freeze()
print(config)
print(args)
main(config, args)

@ -1,220 +0,0 @@
# Copyright (c) 2020 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 time
from collections import defaultdict
import numpy as np
import paddle
from paddle import distributed as dist
from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from paddlespeech.t2s.data import dataset
from paddlespeech.t2s.exps.tacotron2.config import get_cfg_defaults
from paddlespeech.t2s.exps.tacotron2.ljspeech import LJSpeech
from paddlespeech.t2s.exps.tacotron2.ljspeech import LJSpeechCollector
from paddlespeech.t2s.models.tacotron2 import Tacotron2
from paddlespeech.t2s.models.tacotron2 import Tacotron2Loss
from paddlespeech.t2s.training.cli import default_argument_parser
from paddlespeech.t2s.training.experiment import ExperimentBase
from paddlespeech.t2s.utils import display
from paddlespeech.t2s.utils import mp_tools
class Experiment(ExperimentBase):
def compute_losses(self, inputs, outputs):
texts, mel_targets, plens, slens = inputs
mel_outputs = outputs["mel_output"]
mel_outputs_postnet = outputs["mel_outputs_postnet"]
attention_weight = outputs["alignments"]
if self.config.model.use_stop_token:
stop_logits = outputs["stop_logits"]
else:
stop_logits = None
losses = self.criterion(mel_outputs, mel_outputs_postnet, mel_targets,
attention_weight, slens, plens, stop_logits)
return losses
def train_batch(self):
start = time.time()
batch = self.read_batch()
data_loader_time = time.time() - start
self.optimizer.clear_grad()
self.model.train()
texts, mels, text_lens, output_lens = batch
outputs = self.model(texts, text_lens, mels, output_lens)
losses = self.compute_losses(batch, outputs)
loss = losses["loss"]
loss.backward()
self.optimizer.step()
iteration_time = time.time() - start
losses_np = {k: float(v) for k, v in losses.items()}
# logging
msg = "Rank: {}, ".format(dist.get_rank())
msg += "step: {}, ".format(self.iteration)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time,
iteration_time)
msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in losses_np.items())
self.logger.info(msg)
if dist.get_rank() == 0:
for k, v in losses_np.items():
self.visualizer.add_scalar(f"train_loss/{k}", v, self.iteration)
@mp_tools.rank_zero_only
@paddle.no_grad()
def valid(self):
valid_losses = defaultdict(list)
for i, batch in enumerate(self.valid_loader):
texts, mels, text_lens, output_lens = batch
outputs = self.model(texts, text_lens, mels, output_lens)
losses = self.compute_losses(batch, outputs)
for k, v in losses.items():
valid_losses[k].append(float(v))
attention_weights = outputs["alignments"]
self.visualizer.add_figure(
f"valid_sentence_{i}_alignments",
display.plot_alignment(attention_weights[0].numpy().T),
self.iteration)
self.visualizer.add_figure(
f"valid_sentence_{i}_target_spectrogram",
display.plot_spectrogram(mels[0].numpy().T), self.iteration)
self.visualizer.add_figure(
f"valid_sentence_{i}_predicted_spectrogram",
display.plot_spectrogram(outputs['mel_outputs_postnet'][0]
.numpy().T), self.iteration)
# write visual log
valid_losses = {k: np.mean(v) for k, v in valid_losses.items()}
# logging
msg = "Valid: "
msg += "step: {}, ".format(self.iteration)
msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in valid_losses.items())
self.logger.info(msg)
for k, v in valid_losses.items():
self.visualizer.add_scalar(f"valid/{k}", v, self.iteration)
def setup_model(self):
config = self.config
model = Tacotron2(
vocab_size=config.model.vocab_size,
d_mels=config.data.n_mels,
d_encoder=config.model.d_encoder,
encoder_conv_layers=config.model.encoder_conv_layers,
encoder_kernel_size=config.model.encoder_kernel_size,
d_prenet=config.model.d_prenet,
d_attention_rnn=config.model.d_attention_rnn,
d_decoder_rnn=config.model.d_decoder_rnn,
attention_filters=config.model.attention_filters,
attention_kernel_size=config.model.attention_kernel_size,
d_attention=config.model.d_attention,
d_postnet=config.model.d_postnet,
postnet_kernel_size=config.model.postnet_kernel_size,
postnet_conv_layers=config.model.postnet_conv_layers,
reduction_factor=config.model.reduction_factor,
p_encoder_dropout=config.model.p_encoder_dropout,
p_prenet_dropout=config.model.p_prenet_dropout,
p_attention_dropout=config.model.p_attention_dropout,
p_decoder_dropout=config.model.p_decoder_dropout,
p_postnet_dropout=config.model.p_postnet_dropout,
use_stop_token=config.model.use_stop_token)
if self.parallel:
model = paddle.DataParallel(model)
grad_clip = paddle.nn.ClipGradByGlobalNorm(
config.training.grad_clip_thresh)
optimizer = paddle.optimizer.Adam(
learning_rate=config.training.lr,
parameters=model.parameters(),
weight_decay=paddle.regularizer.L2Decay(
config.training.weight_decay),
grad_clip=grad_clip)
criterion = Tacotron2Loss(
use_stop_token_loss=config.model.use_stop_token,
use_guided_attention_loss=config.model.use_guided_attention_loss,
sigma=config.model.guided_attention_loss_sigma)
self.model = model
self.optimizer = optimizer
self.criterion = criterion
def setup_dataloader(self):
args = self.args
config = self.config
ljspeech_dataset = LJSpeech(args.data)
valid_set, train_set = dataset.split(ljspeech_dataset,
config.data.valid_size)
batch_fn = LJSpeechCollector(padding_idx=config.data.padding_idx)
if not self.parallel:
self.train_loader = DataLoader(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True,
collate_fn=batch_fn)
else:
sampler = DistributedBatchSampler(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True)
self.train_loader = DataLoader(
train_set, batch_sampler=sampler, collate_fn=batch_fn)
self.valid_loader = DataLoader(
valid_set,
batch_size=config.data.batch_size,
shuffle=False,
drop_last=False,
collate_fn=batch_fn)
def main_sp(config, args):
exp = Experiment(config, args)
exp.setup()
exp.resume_or_load()
exp.run()
def main(config, args):
if args.ngpu > 1:
dist.spawn(main_sp, args=(config, args), nprocs=args.ngpu)
else:
main_sp(config, args)
if __name__ == "__main__":
config = get_cfg_defaults()
parser = default_argument_parser()
args = parser.parse_args()
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)
config.freeze()
print(config)
print(args)
main(config, args)

@ -130,6 +130,9 @@ def main():
"speech_lengths": item['speech_lengths'],
"speech": str(speech_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"

@ -21,17 +21,43 @@ import soundfile as sf
import yaml
from yacs.config import CfgNode
from paddlespeech.s2t.utils.dynamic_import import dynamic_import
from paddlespeech.t2s.frontend.zh_frontend import Frontend
from paddlespeech.t2s.models.fastspeech2 import FastSpeech2
from paddlespeech.t2s.models.fastspeech2 import FastSpeech2Inference
from paddlespeech.t2s.models.parallel_wavegan import PWGGenerator
from paddlespeech.t2s.models.parallel_wavegan import PWGInference
from paddlespeech.t2s.modules.normalizer import ZScore
from paddlespeech.vector.exps.ge2e.audio_processor import SpeakerVerificationPreprocessor
from paddlespeech.vector.models.lstm_speaker_encoder import LSTMSpeakerEncoder
model_alias = {
# acoustic model
"fastspeech2":
"paddlespeech.t2s.models.fastspeech2:FastSpeech2",
"fastspeech2_inference":
"paddlespeech.t2s.models.fastspeech2:FastSpeech2Inference",
"tacotron2":
"paddlespeech.t2s.models.new_tacotron2:Tacotron2",
"tacotron2_inference":
"paddlespeech.t2s.models.new_tacotron2:Tacotron2Inference",
# voc
"pwgan":
"paddlespeech.t2s.models.parallel_wavegan:PWGGenerator",
"pwgan_inference":
"paddlespeech.t2s.models.parallel_wavegan:PWGInference",
}
def voice_cloning(args):
# Init body.
with open(args.am_config) as f:
am_config = CfgNode(yaml.safe_load(f))
with open(args.voc_config) as f:
voc_config = CfgNode(yaml.safe_load(f))
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(am_config)
print(voc_config)
def voice_cloning(args, fastspeech2_config, pwg_config):
# speaker encoder
p = SpeakerVerificationPreprocessor(
sampling_rate=16000,
@ -57,40 +83,52 @@ def voice_cloning(args, fastspeech2_config, pwg_config):
phn_id = [line.strip().split() for line in f.readlines()]
vocab_size = len(phn_id)
print("vocab_size:", vocab_size)
odim = fastspeech2_config.n_mels
model = FastSpeech2(
idim=vocab_size, odim=odim, **fastspeech2_config["model"])
model.set_state_dict(
paddle.load(args.fastspeech2_checkpoint)["main_params"])
model.eval()
vocoder = PWGGenerator(**pwg_config["generator_params"])
vocoder.set_state_dict(paddle.load(args.pwg_checkpoint)["generator_params"])
vocoder.remove_weight_norm()
vocoder.eval()
print("model done!")
# acoustic model
odim = am_config.n_mels
# model: {model_name}_{dataset}
am_name = args.am[:args.am.rindex('_')]
am_dataset = args.am[args.am.rindex('_') + 1:]
am_class = dynamic_import(am_name, model_alias)
am_inference_class = dynamic_import(am_name + '_inference', model_alias)
if am_name == 'fastspeech2':
am = am_class(
idim=vocab_size, odim=odim, spk_num=None, **am_config["model"])
elif am_name == 'tacotron2':
am = am_class(idim=vocab_size, odim=odim, **am_config["model"])
am.set_state_dict(paddle.load(args.am_ckpt)["main_params"])
am.eval()
am_mu, am_std = np.load(args.am_stat)
am_mu = paddle.to_tensor(am_mu)
am_std = paddle.to_tensor(am_std)
am_normalizer = ZScore(am_mu, am_std)
am_inference = am_inference_class(am_normalizer, am)
am_inference.eval()
print("acoustic model done!")
# vocoder
# model: {model_name}_{dataset}
voc_name = args.voc[:args.voc.rindex('_')]
voc_class = dynamic_import(voc_name, model_alias)
voc_inference_class = dynamic_import(voc_name + '_inference', model_alias)
voc = voc_class(**voc_config["generator_params"])
voc.set_state_dict(paddle.load(args.voc_ckpt)["generator_params"])
voc.remove_weight_norm()
voc.eval()
voc_mu, voc_std = np.load(args.voc_stat)
voc_mu = paddle.to_tensor(voc_mu)
voc_std = paddle.to_tensor(voc_std)
voc_normalizer = ZScore(voc_mu, voc_std)
voc_inference = voc_inference_class(voc_normalizer, voc)
voc_inference.eval()
print("voc done!")
frontend = Frontend(phone_vocab_path=args.phones_dict)
print("frontend done!")
stat = np.load(args.fastspeech2_stat)
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
fastspeech2_normalizer = ZScore(mu, std)
stat = np.load(args.pwg_stat)
mu, std = stat
mu = paddle.to_tensor(mu)
std = paddle.to_tensor(std)
pwg_normalizer = ZScore(mu, std)
fastspeech2_inference = FastSpeech2Inference(fastspeech2_normalizer, model)
fastspeech2_inference.eval()
pwg_inference = PWGInference(pwg_normalizer, vocoder)
pwg_inference.eval()
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
@ -112,24 +150,23 @@ def voice_cloning(args, fastspeech2_config, pwg_config):
# print("spk_emb shape: ", spk_emb.shape)
with paddle.no_grad():
wav = pwg_inference(
fastspeech2_inference(phone_ids, spk_emb=spk_emb))
wav = voc_inference(am_inference(phone_ids, spk_emb=spk_emb))
sf.write(
str(output_dir / (utt_id + ".wav")),
wav.numpy(),
samplerate=fastspeech2_config.fs)
samplerate=am_config.fs)
print(f"{utt_id} done!")
# Randomly generate numbers of 0 ~ 0.2, 256 is the dim of spk_emb
random_spk_emb = np.random.rand(256) * 0.2
random_spk_emb = paddle.to_tensor(random_spk_emb)
utt_id = "random_spk_emb"
with paddle.no_grad():
wav = pwg_inference(fastspeech2_inference(phone_ids, spk_emb=spk_emb))
wav = voc_inference(am_inference(phone_ids, spk_emb=spk_emb))
sf.write(
str(output_dir / (utt_id + ".wav")),
wav.numpy(),
samplerate=fastspeech2_config.fs)
samplerate=am_config.fs)
print(f"{utt_id} done!")
@ -137,32 +174,53 @@ def main():
# parse args and config and redirect to train_sp
parser = argparse.ArgumentParser(description="")
parser.add_argument(
"--fastspeech2-config", type=str, help="fastspeech2 config file.")
parser.add_argument(
"--fastspeech2-checkpoint",
'--am',
type=str,
help="fastspeech2 checkpoint to load.")
default='fastspeech2_csmsc',
choices=['fastspeech2_aishell3', 'tacotron2_aishell3'],
help='Choose acoustic model type of tts task.')
parser.add_argument(
"--fastspeech2-stat",
'--am_config',
type=str,
help="mean and standard deviation used to normalize spectrogram when training fastspeech2."
)
default=None,
help='Config of acoustic model. Use deault config when it is None.')
parser.add_argument(
"--pwg-config", type=str, help="parallel wavegan config file.")
parser.add_argument(
"--pwg-checkpoint",
'--am_ckpt',
type=str,
help="parallel wavegan generator parameters to load.")
default=None,
help='Checkpoint file of acoustic model.')
parser.add_argument(
"--pwg-stat",
"--am_stat",
type=str,
help="mean and standard deviation used to normalize spectrogram when training parallel wavegan."
default=None,
help="mean and standard deviation used to normalize spectrogram when training acoustic model."
)
parser.add_argument(
"--phones-dict",
type=str,
default="phone_id_map.txt",
help="phone vocabulary file.")
# vocoder
parser.add_argument(
'--voc',
type=str,
default='pwgan_csmsc',
choices=['pwgan_aishell3'],
help='Choose vocoder type of tts task.')
parser.add_argument(
'--voc_config',
type=str,
default=None,
help='Config of voc. Use deault config when it is None.')
parser.add_argument(
'--voc_ckpt', type=str, default=None, help='Checkpoint file of voc.')
parser.add_argument(
"--voc_stat",
type=str,
default=None,
help="mean and standard deviation used to normalize spectrogram when training voc."
)
parser.add_argument(
"--text",
type=str,
@ -190,18 +248,7 @@ def main():
else:
print("ngpu should >= 0 !")
with open(args.fastspeech2_config) as f:
fastspeech2_config = CfgNode(yaml.safe_load(f))
with open(args.pwg_config) as f:
pwg_config = CfgNode(yaml.safe_load(f))
print("========Args========")
print(yaml.safe_dump(vars(args)))
print("========Config========")
print(fastspeech2_config)
print(pwg_config)
voice_cloning(args, fastspeech2_config, pwg_config)
voice_cloning(args)
if __name__ == "__main__":

@ -1,13 +0,0 @@
# Copyright (c) 2020 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.

@ -1,13 +0,0 @@
# Copyright (c) 2020 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.

@ -1,89 +0,0 @@
# 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 pickle
from pathlib import Path
import numpy as np
from paddle.io import Dataset
from paddlespeech.t2s.data import batch_spec
from paddlespeech.t2s.data import batch_text_id
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.preprocess_transcription import _phones
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.preprocess_transcription import _tones
from paddlespeech.t2s.frontend import Vocab
voc_phones = Vocab(sorted(list(_phones)))
print("vocab_phones:\n", voc_phones)
voc_tones = Vocab(sorted(list(_tones)))
print("vocab_tones:\n", voc_tones)
class AiShell3(Dataset):
"""Processed AiShell3 dataset."""
def __init__(self, root):
super().__init__()
self.root = Path(root).expanduser()
self.embed_dir = self.root / "embed"
self.mel_dir = self.root / "mel"
with open(self.root / "metadata.pickle", 'rb') as f:
self.records = pickle.load(f)
def __getitem__(self, index):
metadatum = self.records[index]
sentence_id = metadatum["sentence_id"]
speaker_id = sentence_id[:7]
phones = metadatum["phones"]
tones = metadatum["tones"]
phones = np.array(
[voc_phones.lookup(item) for item in phones], dtype=np.int64)
tones = np.array(
[voc_tones.lookup(item) for item in tones], dtype=np.int64)
mel = np.load(str(self.mel_dir / speaker_id / (sentence_id + ".npy")))
embed = np.load(
str(self.embed_dir / speaker_id / (sentence_id + ".npy")))
return phones, tones, mel, embed
def __len__(self):
return len(self.records)
def collate_aishell3_examples(examples):
phones, tones, mel, embed = list(zip(*examples))
text_lengths = np.array([item.shape[0] for item in phones], dtype=np.int64)
spec_lengths = np.array([item.shape[1] for item in mel], dtype=np.int64)
T_dec = np.max(spec_lengths)
stop_tokens = (
np.arange(T_dec) >= np.expand_dims(spec_lengths, -1)).astype(np.float32)
phones, _ = batch_text_id(phones)
tones, _ = batch_text_id(tones)
mel, _ = batch_spec(mel)
mel = np.transpose(mel, (0, 2, 1))
embed = np.stack(embed)
# 7 fields
# (B, T), (B, T), (B, T, C), (B, C), (B,), (B,), (B, T)
return phones, tones, mel, embed, text_lengths, spec_lengths, stop_tokens
if __name__ == "__main__":
dataset = AiShell3("~/datasets/aishell3/train")
example = dataset[0]
examples = [dataset[i] for i in range(10)]
batch = collate_aishell3_examples(examples)
for field in batch:
print(field.shape, field.dtype)

@ -1,42 +0,0 @@
# 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.
from typing import List
from typing import Tuple
from pypinyin import lazy_pinyin
from pypinyin import Style
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.preprocess_transcription import split_syllable
def convert_to_pinyin(text: str) -> List[str]:
"""convert text into list of syllables, other characters that are not chinese, thus
cannot be converted to pinyin are splited.
"""
syllables = lazy_pinyin(
text, style=Style.TONE3, neutral_tone_with_five=True)
return syllables
def convert_sentence(text: str) -> List[Tuple[str]]:
"""convert a sentence into two list: phones and tones"""
syllables = convert_to_pinyin(text)
phones = []
tones = []
for syllable in syllables:
p, t = split_syllable(syllable)
phones.extend(p)
tones.extend(t)
return phones, tones

@ -1,81 +0,0 @@
# Copyright (c) 2020 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 yacs.config import CfgNode as CN
_C = CN()
_C.data = CN(
dict(
batch_size=32, # batch size
valid_size=64, # the first N examples are reserved for validation
sample_rate=22050, # Hz, sample rate
n_fft=1024, # fft frame size
win_length=1024, # window size
hop_length=256, # hop size between ajacent frame
fmax=8000, # Hz, max frequency when converting to mel
fmin=0, # Hz, min frequency when converting to mel
d_mels=80, # mel bands
padding_idx=0, # text embedding's padding index
))
_C.model = CN(
dict(
vocab_size=70,
n_tones=10,
reduction_factor=1, # reduction factor
d_encoder=512, # embedding & encoder's internal size
encoder_conv_layers=3, # number of conv layer in tacotron2 encoder
encoder_kernel_size=5, # kernel size of conv layers in tacotron2 encoder
d_prenet=256, # hidden size of decoder prenet
# hidden size of the first rnn layer in tacotron2 decoder
d_attention_rnn=1024,
# hidden size of the second rnn layer in tacotron2 decoder
d_decoder_rnn=1024,
d_attention=128, # hidden size of decoder location linear layer
attention_filters=32, # number of filter in decoder location conv layer
attention_kernel_size=31, # kernel size of decoder location conv layer
d_postnet=512, # hidden size of decoder postnet
postnet_kernel_size=5, # kernel size of conv layers in postnet
postnet_conv_layers=5, # number of conv layer in decoder postnet
p_encoder_dropout=0.5, # droput probability in encoder
p_prenet_dropout=0.5, # droput probability in decoder prenet
# droput probability of first rnn layer in decoder
p_attention_dropout=0.1,
# droput probability of second rnn layer in decoder
p_decoder_dropout=0.1,
p_postnet_dropout=0.5, # droput probability in decoder postnet
guided_attention_loss_sigma=0.2,
d_global_condition=256,
# whether to use a classifier to predict stop probability
use_stop_token=False,
# whether to use guided attention loss in training
use_guided_attention_loss=True, ))
_C.training = CN(
dict(
lr=1e-3, # learning rate
weight_decay=1e-6, # the coeff of weight decay
grad_clip_thresh=1.0, # the clip norm of grad clip.
valid_interval=1000, # validation
save_interval=1000, # checkpoint
max_iteration=500000, # max iteration to train
))
def get_cfg_defaults():
"""Get a yacs CfgNode object with default values for my_project."""
# Return a clone so that the defaults will not be altered
# This is for the "local variable" use pattern
return _C.clone()

@ -1,95 +0,0 @@
# 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 multiprocessing as mp
from functools import partial
from pathlib import Path
import numpy as np
import tqdm
from paddlespeech.t2s.audio import AudioProcessor
from paddlespeech.t2s.audio.spec_normalizer import LogMagnitude
from paddlespeech.t2s.audio.spec_normalizer import NormalizerBase
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.config import get_cfg_defaults
def extract_mel(fname: Path,
input_dir: Path,
output_dir: Path,
p: AudioProcessor,
n: NormalizerBase):
relative_path = fname.relative_to(input_dir)
out_path = (output_dir / relative_path).with_suffix(".npy")
out_path.parent.mkdir(parents=True, exist_ok=True)
wav = p.read_wav(fname)
mel = p.mel_spectrogram(wav)
mel = n.transform(mel)
np.save(out_path, mel)
def extract_mel_multispeaker(config, input_dir, output_dir, extension=".wav"):
input_dir = Path(input_dir).expanduser()
fnames = list(input_dir.rglob(f"*{extension}"))
output_dir = Path(output_dir).expanduser()
output_dir.mkdir(parents=True, exist_ok=True)
p = AudioProcessor(config.sample_rate, config.n_fft, config.win_length,
config.hop_length, config.d_mels, config.fmin,
config.fmax)
n = LogMagnitude(1e-5)
func = partial(
extract_mel, input_dir=input_dir, output_dir=output_dir, p=p, n=n)
with mp.Pool(16) as pool:
list(
tqdm.tqdm(
pool.imap(func, fnames), total=len(fnames), unit="utterance"))
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Extract mel spectrogram from processed wav in AiShell3 training dataset."
)
parser.add_argument(
"--config",
type=str,
help="yaml config file to overwrite the default config")
parser.add_argument(
"--input",
type=str,
default="~/datasets/aishell3/train/normalized_wav",
help="path of the processed wav folder")
parser.add_argument(
"--output",
type=str,
default="~/datasets/aishell3/train/mel",
help="path of the folder to save mel spectrograms")
parser.add_argument(
"--opts",
nargs=argparse.REMAINDER,
help="options to overwrite --config file and the default config, passing in KEY VALUE pairs"
)
default_config = get_cfg_defaults()
args = parser.parse_args()
if args.config:
default_config.merge_from_file(args.config)
if args.opts:
default_config.merge_from_list(args.opts)
default_config.freeze()
audio_config = default_config.data
extract_mel_multispeaker(audio_config, args.input, args.output)

@ -1,257 +0,0 @@
# 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 pickle
import re
from pathlib import Path
import tqdm
import yaml
zh_pattern = re.compile("[\u4e00-\u9fa5]")
_tones = {'<pad>', '<s>', '</s>', '0', '1', '2', '3', '4', '5'}
_pauses = {'%', '$'}
_initials = {
'b',
'p',
'm',
'f',
'd',
't',
'n',
'l',
'g',
'k',
'h',
'j',
'q',
'x',
'zh',
'ch',
'sh',
'r',
'z',
'c',
's',
}
_finals = {
'ii',
'iii',
'a',
'o',
'e',
'ea',
'ai',
'ei',
'ao',
'ou',
'an',
'en',
'ang',
'eng',
'er',
'i',
'ia',
'io',
'ie',
'iai',
'iao',
'iou',
'ian',
'ien',
'iang',
'ieng',
'u',
'ua',
'uo',
'uai',
'uei',
'uan',
'uen',
'uang',
'ueng',
'v',
've',
'van',
'ven',
'veng',
}
_ernized_symbol = {'&r'}
_specials = {'<pad>', '<unk>', '<s>', '</s>'}
_phones = _initials | _finals | _ernized_symbol | _specials | _pauses
def is_zh(word):
global zh_pattern
match = zh_pattern.search(word)
return match is not None
def ernized(syllable):
return syllable[:2] != "er" and syllable[-2] == 'r'
def convert(syllable):
# expansion of o -> uo
syllable = re.sub(r"([bpmf])o$", r"\1uo", syllable)
# syllable = syllable.replace("bo", "buo").replace("po", "puo").replace("mo", "muo").replace("fo", "fuo")
# expansion for iong, ong
syllable = syllable.replace("iong", "veng").replace("ong", "ueng")
# expansion for ing, in
syllable = syllable.replace("ing", "ieng").replace("in", "ien")
# expansion for un, ui, iu
syllable = syllable.replace("un", "uen").replace("ui",
"uei").replace("iu", "iou")
# rule for variants of i
syllable = syllable.replace("zi", "zii").replace("ci", "cii").replace("si", "sii")\
.replace("zhi", "zhiii").replace("chi", "chiii").replace("shi", "shiii")\
.replace("ri", "riii")
# rule for y preceding i, u
syllable = syllable.replace("yi", "i").replace("yu", "v").replace("y", "i")
# rule for w
syllable = syllable.replace("wu", "u").replace("w", "u")
# rule for v following j, q, x
syllable = syllable.replace("ju", "jv").replace("qu",
"qv").replace("xu", "xv")
return syllable
def split_syllable(syllable: str):
"""Split a syllable in pinyin into a list of phones and a list of tones.
Initials have no tone, represented by '0', while finals have tones from
'1,2,3,4,5'.
e.g.
zhang -> ['zh', 'ang'], ['0', '1']
"""
if syllable in _pauses:
# syllable, tone
return [syllable], ['0']
tone = syllable[-1]
syllable = convert(syllable[:-1])
phones = []
tones = []
global _initials
if syllable[:2] in _initials:
phones.append(syllable[:2])
tones.append('0')
phones.append(syllable[2:])
tones.append(tone)
elif syllable[0] in _initials:
phones.append(syllable[0])
tones.append('0')
phones.append(syllable[1:])
tones.append(tone)
else:
phones.append(syllable)
tones.append(tone)
return phones, tones
def load_aishell3_transcription(line: str):
sentence_id, pinyin, text = line.strip().split("|")
syllables = pinyin.strip().split()
results = []
for syllable in syllables:
if syllable in _pauses:
results.append(syllable)
elif not ernized(syllable):
results.append(syllable)
else:
results.append(syllable[:-2] + syllable[-1])
results.append('&r5')
phones = []
tones = []
for syllable in results:
p, t = split_syllable(syllable)
phones.extend(p)
tones.extend(t)
for p in phones:
assert p in _phones, p
return {
"sentence_id": sentence_id,
"text": text,
"syllables": results,
"phones": phones,
"tones": tones
}
def process_aishell3(dataset_root, output_dir):
dataset_root = Path(dataset_root).expanduser()
output_dir = Path(output_dir).expanduser()
output_dir.mkdir(parents=True, exist_ok=True)
prosody_label_path = dataset_root / "label_train-set.txt"
with open(prosody_label_path, 'rt') as f:
lines = [line.strip() for line in f]
records = lines[5:]
processed_records = []
for record in tqdm.tqdm(records):
new_record = load_aishell3_transcription(record)
processed_records.append(new_record)
print(new_record)
with open(output_dir / "metadata.pickle", 'wb') as f:
pickle.dump(processed_records, f)
with open(output_dir / "metadata.yaml", 'wt', encoding="utf-8") as f:
yaml.safe_dump(
processed_records, f, default_flow_style=None, allow_unicode=True)
print("metadata done!")
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Preprocess transcription of AiShell3 and save them in a compact file(yaml and pickle)."
)
parser.add_argument(
"--input",
type=str,
default="~/datasets/aishell3/train",
help="path of the training dataset,(contains a label_train-set.txt).")
parser.add_argument(
"--output",
type=str,
help="the directory to save the processed transcription."
"If not provided, it would be the same as the input.")
args = parser.parse_args()
if args.output is None:
args.output = args.input
process_aishell3(args.input, args.output)

@ -1,94 +0,0 @@
# 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 functools import partial
from multiprocessing import Pool
from pathlib import Path
import librosa
import numpy as np
import soundfile as sf
from praatio import textgrid
from tqdm import tqdm
def get_valid_part(fpath):
f = textgrid.openTextgrid(fpath, includeEmptyIntervals=True)
start = 0
phone_entry_list = f.tierDict['phones'].entryList
first_entry = phone_entry_list[0]
if first_entry.label == "sil":
start = first_entry.end
last_entry = phone_entry_list[-1]
if last_entry.label == "sp":
end = last_entry.start
else:
end = last_entry.end
return start, end
def process_utterance(fpath, source_dir, target_dir, alignment_dir):
rel_path = fpath.relative_to(source_dir)
opath = target_dir / rel_path
apath = (alignment_dir / rel_path).with_suffix(".TextGrid")
opath.parent.mkdir(parents=True, exist_ok=True)
start, end = get_valid_part(apath)
wav, _ = librosa.load(fpath, sr=22050, offset=start, duration=end - start)
normalized_wav = wav / np.max(wav) * 0.999
sf.write(opath, normalized_wav, samplerate=22050, subtype='PCM_16')
# print(f"{fpath} => {opath}")
def preprocess_aishell3(source_dir, target_dir, alignment_dir):
source_dir = Path(source_dir).expanduser()
target_dir = Path(target_dir).expanduser()
alignment_dir = Path(alignment_dir).expanduser()
wav_paths = list(source_dir.rglob("*.wav"))
print(f"there are {len(wav_paths)} audio files in total")
fx = partial(
process_utterance,
source_dir=source_dir,
target_dir=target_dir,
alignment_dir=alignment_dir)
with Pool(16) as p:
list(
tqdm(p.imap(fx, wav_paths), total=len(wav_paths), unit="utterance"))
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Process audio in AiShell3, trim silence according to the alignment "
"files generated by MFA, and normalize volume by peak.")
parser.add_argument(
"--input",
type=str,
default="~/datasets/aishell3/train/wav",
help="path of the original audio folder in aishell3.")
parser.add_argument(
"--output",
type=str,
default="~/datasets/aishell3/train/normalized_wav",
help="path of the folder to save the processed audio files.")
parser.add_argument(
"--alignment",
type=str,
default="~/datasets/aishell3/train/alignment",
help="path of the alignment files.")
args = parser.parse_args()
preprocess_aishell3(args.input, args.output, args.alignment)

@ -1,263 +0,0 @@
# Copyright (c) 2020 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 time
from collections import defaultdict
from pathlib import Path
import numpy as np
import paddle
from matplotlib import pyplot as plt
from paddle import distributed as dist
from paddle.io import DataLoader
from paddle.io import DistributedBatchSampler
from paddlespeech.t2s.data import dataset
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.aishell3 import AiShell3
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.aishell3 import collate_aishell3_examples
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.config import get_cfg_defaults
from paddlespeech.t2s.models.tacotron2 import Tacotron2
from paddlespeech.t2s.models.tacotron2 import Tacotron2Loss
from paddlespeech.t2s.training.cli import default_argument_parser
from paddlespeech.t2s.training.experiment import ExperimentBase
from paddlespeech.t2s.utils import display
from paddlespeech.t2s.utils import mp_tools
class Experiment(ExperimentBase):
def compute_losses(self, inputs, outputs):
texts, tones, mel_targets, utterance_embeds, text_lens, output_lens, stop_tokens = inputs
mel_outputs = outputs["mel_output"]
mel_outputs_postnet = outputs["mel_outputs_postnet"]
alignments = outputs["alignments"]
losses = self.criterion(mel_outputs, mel_outputs_postnet, mel_targets,
alignments, output_lens, text_lens)
return losses
def train_batch(self):
start = time.time()
batch = self.read_batch()
data_loader_time = time.time() - start
self.optimizer.clear_grad()
self.model.train()
texts, tones, mels, utterance_embeds, text_lens, output_lens, stop_tokens = batch
outputs = self.model(
texts,
text_lens,
mels,
output_lens,
tones=tones,
global_condition=utterance_embeds)
losses = self.compute_losses(batch, outputs)
loss = losses["loss"]
loss.backward()
self.optimizer.step()
iteration_time = time.time() - start
losses_np = {k: float(v) for k, v in losses.items()}
# logging
msg = "Rank: {}, ".format(dist.get_rank())
msg += "step: {}, ".format(self.iteration)
msg += "time: {:>.3f}s/{:>.3f}s, ".format(data_loader_time,
iteration_time)
msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in losses_np.items())
self.logger.info(msg)
if dist.get_rank() == 0:
for key, value in losses_np.items():
self.visualizer.add_scalar(f"train_loss/{key}", value,
self.iteration)
@mp_tools.rank_zero_only
@paddle.no_grad()
def valid(self):
valid_losses = defaultdict(list)
for i, batch in enumerate(self.valid_loader):
texts, tones, mels, utterance_embeds, text_lens, output_lens, stop_tokens = batch
outputs = self.model(
texts,
text_lens,
mels,
output_lens,
tones=tones,
global_condition=utterance_embeds)
losses = self.compute_losses(batch, outputs)
for key, value in losses.items():
valid_losses[key].append(float(value))
attention_weights = outputs["alignments"]
self.visualizer.add_figure(
f"valid_sentence_{i}_alignments",
display.plot_alignment(attention_weights[0].numpy().T),
self.iteration)
self.visualizer.add_figure(
f"valid_sentence_{i}_target_spectrogram",
display.plot_spectrogram(mels[0].numpy().T), self.iteration)
mel_pred = outputs['mel_outputs_postnet']
self.visualizer.add_figure(
f"valid_sentence_{i}_predicted_spectrogram",
display.plot_spectrogram(mel_pred[0].numpy().T), self.iteration)
# write visual log
valid_losses = {k: np.mean(v) for k, v in valid_losses.items()}
# logging
msg = "Valid: "
msg += "step: {}, ".format(self.iteration)
msg += ', '.join('{}: {:>.6f}'.format(k, v)
for k, v in valid_losses.items())
self.logger.info(msg)
for key, value in valid_losses.items():
self.visualizer.add_scalar(f"valid/{key}", value, self.iteration)
@mp_tools.rank_zero_only
@paddle.no_grad()
def eval(self):
"""Evaluation of Tacotron2 in autoregressive manner."""
self.model.eval()
mel_dir = Path(self.output_dir / ("eval_{}".format(self.iteration)))
mel_dir.mkdir(parents=True, exist_ok=True)
for i, batch in enumerate(self.test_loader):
texts, tones, mels, utterance_embeds, *_ = batch
outputs = self.model.infer(
texts, tones=tones, global_condition=utterance_embeds)
display.plot_alignment(outputs["alignments"][0].numpy().T)
plt.savefig(mel_dir / f"sentence_{i}.png")
plt.close()
np.save(mel_dir / f"sentence_{i}",
outputs["mel_outputs_postnet"][0].numpy().T)
print(f"sentence_{i}")
def setup_model(self):
config = self.config
model = Tacotron2(
vocab_size=config.model.vocab_size,
n_tones=config.model.n_tones,
d_mels=config.data.d_mels,
d_encoder=config.model.d_encoder,
encoder_conv_layers=config.model.encoder_conv_layers,
encoder_kernel_size=config.model.encoder_kernel_size,
d_prenet=config.model.d_prenet,
d_attention_rnn=config.model.d_attention_rnn,
d_decoder_rnn=config.model.d_decoder_rnn,
attention_filters=config.model.attention_filters,
attention_kernel_size=config.model.attention_kernel_size,
d_attention=config.model.d_attention,
d_postnet=config.model.d_postnet,
postnet_kernel_size=config.model.postnet_kernel_size,
postnet_conv_layers=config.model.postnet_conv_layers,
reduction_factor=config.model.reduction_factor,
p_encoder_dropout=config.model.p_encoder_dropout,
p_prenet_dropout=config.model.p_prenet_dropout,
p_attention_dropout=config.model.p_attention_dropout,
p_decoder_dropout=config.model.p_decoder_dropout,
p_postnet_dropout=config.model.p_postnet_dropout,
d_global_condition=config.model.d_global_condition,
use_stop_token=config.model.use_stop_token, )
if self.parallel:
model = paddle.DataParallel(model)
grad_clip = paddle.nn.ClipGradByGlobalNorm(
config.training.grad_clip_thresh)
optimizer = paddle.optimizer.Adam(
learning_rate=config.training.lr,
parameters=model.parameters(),
weight_decay=paddle.regularizer.L2Decay(
config.training.weight_decay),
grad_clip=grad_clip)
criterion = Tacotron2Loss(
use_stop_token_loss=config.model.use_stop_token,
use_guided_attention_loss=config.model.use_guided_attention_loss,
sigma=config.model.guided_attention_loss_sigma)
self.model = model
self.optimizer = optimizer
self.criterion = criterion
def setup_dataloader(self):
args = self.args
config = self.config
aishell3_dataset = AiShell3(args.data)
valid_set, train_set = dataset.split(aishell3_dataset,
config.data.valid_size)
batch_fn = collate_aishell3_examples
if not self.parallel:
self.train_loader = DataLoader(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True,
collate_fn=batch_fn)
else:
sampler = DistributedBatchSampler(
train_set,
batch_size=config.data.batch_size,
shuffle=True,
drop_last=True)
self.train_loader = DataLoader(
train_set, batch_sampler=sampler, collate_fn=batch_fn)
self.valid_loader = DataLoader(
valid_set,
batch_size=config.data.batch_size,
shuffle=False,
drop_last=False,
collate_fn=batch_fn)
self.test_loader = DataLoader(
valid_set,
batch_size=1,
shuffle=False,
drop_last=False,
collate_fn=batch_fn)
def main_sp(config, args):
exp = Experiment(config, args)
exp.setup()
exp.resume_or_load()
if not args.test:
exp.run()
else:
exp.eval()
def main(config, args):
if args.ngpu > 1:
dist.spawn(main_sp, args=(config, args), nprocs=args.ngpu)
else:
main_sp(config, args)
if __name__ == "__main__":
config = get_cfg_defaults()
parser = default_argument_parser()
parser.add_argument("--test", action="store_true")
args = parser.parse_args()
if args.config:
config.merge_from_file(args.config)
if args.opts:
config.merge_from_list(args.opts)
config.freeze()
print(config)
print(args)
main(config, args)

@ -1,166 +0,0 @@
# 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 pathlib import Path
import numpy as np
import paddle
import soundfile as sf
from matplotlib import pyplot as plt
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.aishell3 import voc_phones
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.aishell3 import voc_tones
from paddlespeech.t2s.exps.voice_cloning.tacotron2_ge2e.chinese_g2p import convert_sentence
from paddlespeech.t2s.models.tacotron2 import Tacotron2
from paddlespeech.t2s.models.waveflow import ConditionalWaveFlow
from paddlespeech.t2s.utils import display
from paddlespeech.vector.exps.ge2e.audio_processor import SpeakerVerificationPreprocessor
from paddlespeech.vector.models.lstm_speaker_encoder import LSTMSpeakerEncoder
def voice_cloning(args):
# speaker encoder
p = SpeakerVerificationPreprocessor(
sampling_rate=16000,
audio_norm_target_dBFS=-30,
vad_window_length=30,
vad_moving_average_width=8,
vad_max_silence_length=6,
mel_window_length=25,
mel_window_step=10,
n_mels=40,
partial_n_frames=160,
min_pad_coverage=0.75,
partial_overlap_ratio=0.5)
print("Audio Processor Done!")
speaker_encoder = LSTMSpeakerEncoder(
n_mels=40, num_layers=3, hidden_size=256, output_size=256)
speaker_encoder.set_state_dict(paddle.load(args.ge2e_params_path))
speaker_encoder.eval()
print("GE2E Done!")
synthesizer = Tacotron2(
vocab_size=68,
n_tones=10,
d_mels=80,
d_encoder=512,
encoder_conv_layers=3,
encoder_kernel_size=5,
d_prenet=256,
d_attention_rnn=1024,
d_decoder_rnn=1024,
attention_filters=32,
attention_kernel_size=31,
d_attention=128,
d_postnet=512,
postnet_kernel_size=5,
postnet_conv_layers=5,
reduction_factor=1,
p_encoder_dropout=0.5,
p_prenet_dropout=0.5,
p_attention_dropout=0.1,
p_decoder_dropout=0.1,
p_postnet_dropout=0.5,
d_global_condition=256,
use_stop_token=False, )
synthesizer.set_state_dict(paddle.load(args.tacotron2_params_path))
synthesizer.eval()
print("Tacotron2 Done!")
# vocoder
vocoder = ConditionalWaveFlow(
upsample_factors=[16, 16],
n_flows=8,
n_layers=8,
n_group=16,
channels=128,
n_mels=80,
kernel_size=[3, 3])
vocoder.set_state_dict(paddle.load(args.waveflow_params_path))
vocoder.eval()
print("WaveFlow Done!")
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
input_dir = Path(args.input_dir)
# 因为 AISHELL-3 数据集中使用 % 和 $ 表示韵律词和韵律短语的边界,它们大约对应着较短和较长的停顿,在文本中可以使用 % 和 $ 来调节韵律。
# 值得的注意的是,句子的有效字符集仅包含汉字和 %, $, 因此输入的句子只能包含这些字符。
sentence = "每当你觉得%想要批评什么人的时候$你切要记着%这个世界上的人%并非都具备你禀有的条件$"
phones, tones = convert_sentence(sentence)
phones = np.array(
[voc_phones.lookup(item) for item in phones], dtype=np.int64)
tones = np.array([voc_tones.lookup(item) for item in tones], dtype=np.int64)
phones = paddle.to_tensor(phones).unsqueeze(0)
tones = paddle.to_tensor(tones).unsqueeze(0)
for name in os.listdir(input_dir):
utt_id = name.split(".")[0]
ref_audio_path = input_dir / name
mel_sequences = p.extract_mel_partials(p.preprocess_wav(ref_audio_path))
print("mel_sequences: ", mel_sequences.shape)
with paddle.no_grad():
embed = speaker_encoder.embed_utterance(
paddle.to_tensor(mel_sequences))
print("embed shape: ", embed.shape)
utterance_embeds = paddle.unsqueeze(embed, 0)
outputs = synthesizer.infer(
phones, tones=tones, global_condition=utterance_embeds)
mel_input = paddle.transpose(outputs["mel_outputs_postnet"], [0, 2, 1])
alignment = outputs["alignments"][0].numpy().T
display.plot_alignment(alignment)
plt.savefig(str(output_dir / (utt_id + ".png")))
with paddle.no_grad():
wav = vocoder.infer(mel_input)
wav = wav.numpy()[0]
sf.write(str(output_dir / (utt_id + ".wav")), wav, samplerate=22050)
def main():
# parse args and config and redirect to train_sp
parser = argparse.ArgumentParser(description="")
parser.add_argument(
"--ge2e_params_path", type=str, help="ge2e params path.")
parser.add_argument(
"--tacotron2_params_path", type=str, help="tacotron2 params path.")
parser.add_argument(
"--waveflow_params_path", type=str, help="waveflow params path.")
parser.add_argument(
"--ngpu", type=int, default=1, help="if ngpu=0, use cpu.")
parser.add_argument(
"--input-dir",
type=str,
help="input dir of *.wav, the sample rate will be resample to 16k.")
parser.add_argument("--output-dir", type=str, help="output dir.")
args = parser.parse_args()
if args.ngpu == 0:
paddle.set_device("cpu")
elif args.ngpu > 0:
paddle.set_device("gpu")
else:
print("ngpu should >= 0 !")
voice_cloning(args)
if __name__ == "__main__":
main()

@ -17,6 +17,5 @@ from .melgan import *
from .new_tacotron2 import *
from .parallel_wavegan import *
from .speedyspeech import *
from .tacotron2 import *
from .transformer_tts import *
from .waveflow import *

@ -479,7 +479,7 @@ class Tacotron2(nn.Layer):
elif self.spk_embed_integration_type == "concat":
# concat hidden states with spk embeds
spk_emb = F.normalize(spk_emb).unsqueeze(1).expand(
-1, paddle.shape(hs)[1], -1)
shape=[-1, paddle.shape(hs)[1], -1])
hs = paddle.concat([hs, spk_emb], axis=-1)
else:
raise NotImplementedError("support only add or concat.")

File diff suppressed because it is too large Load Diff

@ -16,3 +16,7 @@ from . import display
from . import layer_tools
from . import mp_tools
from . import scheduler
def str2bool(str):
return True if str.lower() == 'true' else False

@ -123,9 +123,3 @@ class Collate(object):
frame_clips = [self.random_crop(mel) for mel in examples]
batced_clips = np.stack(frame_clips)
return batced_clips
if __name__ == "__main__":
mydataset = MultiSpeakerMelDataset(
Path("/home/chenfeiyu/datasets/SV2TTS/encoder"))
print(mydataset.get_example_by_index(0, 10))

@ -22,6 +22,7 @@ from sklearn.preprocessing import StandardScaler
from tqdm import tqdm
from paddlespeech.t2s.datasets.data_table import DataTable
from paddlespeech.t2s.utils import str2bool
def main():
@ -41,9 +42,6 @@ def main():
help="path to save statistics. if not provided, "
"stats will be saved in the above root directory with name stats.npy")
def str2bool(str):
return True if str.lower() == 'true' else False
parser.add_argument(
"--use-relative-path",
type=str2bool,

Loading…
Cancel
Save