You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
PaddleSpeech/examples/vctk/fastspeech2/baker
Hui Zhang b079577e08
merge parakeet repo into deepspeech
3 years ago
..
conf merge parakeet repo into deepspeech 3 years ago
README.md merge parakeet repo into deepspeech 3 years ago
preprocess.sh merge parakeet repo into deepspeech 3 years ago
run.sh merge parakeet repo into deepspeech 3 years ago
simple.lexicon merge parakeet repo into deepspeech 3 years ago
synthesize.sh merge parakeet repo into deepspeech 3 years ago
synthesize_e2e.py merge parakeet repo into deepspeech 3 years ago
synthesize_e2e.sh merge parakeet repo into deepspeech 3 years ago

README.md

FastSpeech2 with CSMSC

This example contains code used to train a Fastspeech2 model with Chinese Standard Mandarin Speech Copus.

Dataset

Download and Extract the datasaet

Download CSMSC from it's Official Website.

Get MFA result of CSMSC and Extract it

We use MFA to get durations for fastspeech2. You can download from here baker_alignment_tone.tar.gz, or train your own MFA model reference to use_mfa example of our repo.

Preprocess the dataset

Assume the path to the dataset is ~/datasets/BZNSYP. Assume the path to the MFA result of CSMSC is ./baker_alignment_tone. Run the command below to preprocess the dataset.

./preprocess.sh

When it is done. A dump folder is created in the current directory. The structure of the dump folder is listed below.

dump
├── dev
│   ├── norm
│   └── raw
├── phone_id_map.txt
├── speaker_id_map.txt
├── test
│   ├── 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 sub folder. The raw folder contains speech、pitch and energy features of each utterances, 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 which contains phones, text_lengths, speech_lengths, durations, path of speech features, path of pitch features, path of energy features, speaker and id of each utterance.

Train the model

./run.sh calls ../train.py.

./run.sh

Here's the complete help message.

usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
                [--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
                [--device DEVICE] [--nprocs NPROCS] [--verbose VERBOSE]
                [--phones-dict PHONES_DICT] [--speaker-dict SPEAKER_DICT]

Train a FastSpeech2 model.

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       fastspeech2 config file.
  --train-metadata TRAIN_METADATA
                        training data.
  --dev-metadata DEV_METADATA
                        dev data.
  --output-dir OUTPUT_DIR
                        output dir.
  --device DEVICE       device type to use.
  --nprocs NPROCS       number of processes.
  --verbose VERBOSE     verbose.
  --phones-dict PHONES_DICT
                        phone vocabulary file.
  --speaker-dict SPEAKER_DICT
                        speaker id map file for multiple speaker model.
  1. --config is a config file in yaml format to overwrite the default config, which can be found at conf/default.yaml.
  2. --train-metadata and --dev-metadata should be the metadata file in the normalized subfolder of train and dev in the dump folder.
  3. --output-dir is the directory to save the results of the experiment. Checkpoints are save in checkpoints/ inside this directory.
  4. --device is the type of the device to run the experiment, 'cpu' or 'gpu' are supported.
  5. --nprocs is the number of processes to run in parallel, note that nprocs > 1 is only supported when --device is 'gpu'.
  6. --phones-dict is the path of the phone vocabulary file.

Pretrained Model

Pretrained FastSpeech2 model with no silence in the edge of audios. fastspeech2_nosil_baker_ckpt_0.4.zip

FastSpeech2 checkpoint contains files listed below.

fastspeech2_nosil_baker_ckpt_0.4
├── default.yaml            # default config used to train fastspeech2
├── phone_id_map.txt        # phone vocabulary file when training fastspeech2
├── snapshot_iter_76000.pdz # model parameters and optimizer states
└── speech_stats.npy        # statistics used to normalize spectrogram when training fastspeech2

Synthesize

We use parallel wavegan as the neural vocoder. Download pretrained parallel wavegan model from pwg_baker_ckpt_0.4.zip and unzip it.

unzip pwg_baker_ckpt_0.4.zip

Parallel WaveGAN checkpoint contains files listed below.

pwg_baker_ckpt_0.4
├── pwg_default.yaml               # default config used to train parallel wavegan
├── pwg_snapshot_iter_400000.pdz   # model parameters of parallel wavegan
└── pwg_stats.npy                  # statistics used to normalize spectrogram when training parallel wavegan

synthesize.sh calls ../synthesize.py, which can synthesize waveform from metadata.jsonl.

./synthesize.sh
usage: synthesize.py [-h] [--fastspeech2-config FASTSPEECH2_CONFIG]
                     [--fastspeech2-checkpoint FASTSPEECH2_CHECKPOINT]
                     [--fastspeech2-stat FASTSPEECH2_STAT]
                     [--pwg-config PWG_CONFIG]
                     [--pwg-checkpoint PWG_CHECKPOINT] [--pwg-stat PWG_STAT]
                     [--phones-dict PHONES_DICT] [--speaker-dict SPEAKER_DICT]
                     [--test-metadata TEST_METADATA] [--output-dir OUTPUT_DIR]
                     [--device DEVICE] [--verbose VERBOSE]

Synthesize with fastspeech2 & parallel wavegan.

optional arguments:
  -h, --help            show this help message and exit
  --fastspeech2-config FASTSPEECH2_CONFIG
                        fastspeech2 config file.
  --fastspeech2-checkpoint FASTSPEECH2_CHECKPOINT
                        fastspeech2 checkpoint to load.
  --fastspeech2-stat FASTSPEECH2_STAT
                        mean and standard deviation used to normalize
                        spectrogram when training fastspeech2.
  --pwg-config PWG_CONFIG
                        parallel wavegan config file.
  --pwg-checkpoint PWG_CHECKPOINT
                        parallel wavegan generator parameters to load.
  --pwg-stat PWG_STAT   mean and standard deviation used to normalize
                        spectrogram when training parallel wavegan.
  --phones-dict PHONES_DICT
                        phone vocabulary file.
  --speaker-dict SPEAKER_DICT
                        speaker id map file for multiple speaker model.
  --test-metadata TEST_METADATA
                        test metadata.
  --output-dir OUTPUT_DIR
                        output dir.
  --device DEVICE       device type to use.
  --verbose VERBOSE     verbose.

synthesize_e2e.sh calls synthesize_e2e.py, which can synthesize waveform from text file.

./synthesize_e2e.sh
usage: synthesize_e2e.py [-h] [--fastspeech2-config FASTSPEECH2_CONFIG]
                         [--fastspeech2-checkpoint FASTSPEECH2_CHECKPOINT]
                         [--fastspeech2-stat FASTSPEECH2_STAT]
                         [--pwg-config PWG_CONFIG]
                         [--pwg-checkpoint PWG_CHECKPOINT]
                         [--pwg-stat PWG_STAT] [--phones-dict PHONES_DICT]
                         [--text TEXT] [--output-dir OUTPUT_DIR]
                         [--device DEVICE] [--verbose VERBOSE]

Synthesize with fastspeech2 & parallel wavegan.

optional arguments:
  -h, --help            show this help message and exit
  --fastspeech2-config FASTSPEECH2_CONFIG
                        fastspeech2 config file.
  --fastspeech2-checkpoint FASTSPEECH2_CHECKPOINT
                        fastspeech2 checkpoint to load.
  --fastspeech2-stat FASTSPEECH2_STAT
                        mean and standard deviation used to normalize
                        spectrogram when training fastspeech2.
  --pwg-config PWG_CONFIG
                        parallel wavegan config file.
  --pwg-checkpoint PWG_CHECKPOINT
                        parallel wavegan generator parameters to load.
  --pwg-stat PWG_STAT   mean and standard deviation used to normalize
                        spectrogram when training parallel wavegan.
  --phones-dict PHONES_DICT
                        phone vocabulary file.
  --text TEXT           text to synthesize, a 'utt_id sentence' pair per line.
  --output-dir OUTPUT_DIR
                        output dir.
  --device DEVICE       device type to use.
  --verbose VERBOSE     verbose.
  1. --fastspeech2-config, --fastspeech2-checkpoint, --fastspeech2-stat and --phones-dict are arguments for fastspeech2, which correspond to the 4 files in the fastspeech2 pretrained model.
  2. --pwg-config, --pwg-checkpoint, --pwg-stat are arguments for parallel wavegan, which correspond to the 3 files in the parallel wavegan pretrained model.
  3. --test-metadata should be the metadata file in the normalized subfolder of test in the dump folder.
  4. --text is the text file, which contains sentences to synthesize.
  5. --output-dir is the directory to save synthesized audio files.
  6. --device is the type of device to run synthesis, 'cpu' and 'gpu' are supported. 'gpu' is recommended for faster synthesis.

You can use the following scripts to synthesize for ../sentences.txt using pretrained fastspeech2 and parallel wavegan models.

FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 synthesize_e2e.py \
  --fastspeech2-config=fastspeech2_nosil_baker_ckpt_0.4/default.yaml \
  --fastspeech2-checkpoint=fastspeech2_nosil_baker_ckpt_0.4/snapshot_iter_76000.pdz \
  --fastspeech2-stat=fastspeech2_nosil_baker_ckpt_0.4/speech_stats.npy \
  --pwg-config=pwg_baker_ckpt_0.4/pwg_default.yaml \
  --pwg-checkpoint=pwg_baker_ckpt_0.4/pwg_snapshot_iter_400000.pdz \
  --pwg-stat=pwg_baker_ckpt_0.4/pwg_stats.npy \
  --text=../sentences.txt \
  --output-dir=exp/default/test_e2e \
  --device="gpu" \
  --phones-dict=fastspeech2_nosil_baker_ckpt_0.4/phone_id_map.txt