PaddleSpeech/examples/csmsc/voc3/README.md

Multi Band MelGAN with CSMSC

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

Dataset

Download and Extract

Download CSMSC from the official website and extract it to ~/datasets. Then the dataset is in directory ~/datasets/BZNSYP.

Get MFA Result and Extract

We use MFA results to cut silence in the edge of audio. You can download from here baker_alignment_tone.tar.gz, or train your own MFA model reference to mfa example of our repo.

Get Started

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

1. source path.
2. preprocess the dataset.
3. train the model.
4. synthesize wavs.
   • synthesize waveform from metadata.jsonl.

./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.

./run.sh --stage 0 --stop-stage 0

Data Preprocessing

./local/preprocess.sh ${conf_path}

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

dump
├── dev
│   ├── norm
│   └── raw
├── test
│   ├── norm
│   └── raw
└── train
    ├── norm
    ├── raw
    └── feats_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 log magnitude of mel spectrogram of each utterances, while the norm folder contains normalized spectrogram. The statistics used to normalize the spectrogram is computed from the training set, which is located in dump/train/feats_stats.npy.

Also there is a metadata.jsonl in each subfolder. It is a table-like file which contains id and paths to spectrogam of each utterance.

Model Training

CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}

./local/train.sh calls ${BIN_DIR}/train.py. Here's the complete help message.

usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
                [--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
                [--ngpu NGPU] [--verbose VERBOSE]

Train a Multi-Band MelGAN model.

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       config file to overwrite default config.
  --train-metadata TRAIN_METADATA
                        training data.
  --dev-metadata DEV_METADATA
                        dev data.
  --output-dir OUTPUT_DIR
                        output dir.
  --ngpu NGPU           if ngpu == 0, use cpu.
  --verbose VERBOSE     verbose.

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. --ngpu is the number of gpus to use, if ngpu == 0, use cpu.

Synthesizing

./local/synthesize.sh calls ${BIN_DIR}/synthesize.py, which can synthesize waveform from metadata.jsonl.

CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}

usage: synthesize.py [-h] [--config CONFIG] [--checkpoint CHECKPOINT]
                     [--test-metadata TEST_METADATA] [--output-dir OUTPUT_DIR]
                     [--ngpu NGPU] [--verbose VERBOSE]

Synthesize with multi band melgan.

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       multi band melgan config file.
  --checkpoint CHECKPOINT
                        snapshot to load.
  --test-metadata TEST_METADATA
                        dev data.
  --output-dir OUTPUT_DIR
                        output dir.
  --ngpu NGPU           if ngpu == 0, use cpu.
  --verbose VERBOSE     verbose.

1. --config multi band melgan config file. You should use the same config with which the model is trained.
2. --checkpoint is the checkpoint to load. Pick one of the checkpoints from checkpoints inside the training output directory.
3. --test-metadata is the metadata of the test dataset. Use the metadata.jsonl in the dev/norm subfolder from the processed directory.
4. --output-dir is the directory to save the synthesized audio files.
5. --ngpu is the number of gpus to use, if ngpu == 0, use cpu.

Fine-tuning

Since there are no noise in the input of Multi Band MelGAN, the audio quality is not so good (see espnet issue), we refer to the method proposed in HiFiGAN, finetune Multi Band MelGAN with the predicted mel-spectrogram from FastSpeech2.

The length of mel-spectrograms should align with the length of wavs, so we should generate mels using ground truth alignment.

But since we are fine-tuning, we should use the statistics computed during training step.

You should first download pretrained FastSpeech2 model from fastspeech2_nosil_baker_ckpt_0.4.zip and unzip it.

Assume the path to the dump-dir of training step is dump. Assume the path to the duration result of CSMSC is durations.txt (generated during training step's preprocessing). Assume the path to the pretrained FastSpeech2 model is fastspeech2_nosil_baker_ckpt_0.4.
The finetune.sh can

1. source path.
2. generate ground truth alignment mels.
3. link *_wave.npy from dump to dump_finetune (because we only use new mels, the wavs are the ones used during train step) .
4. copy features' stats from dump to dump_finetune.
5. normalize the ground truth alignment mels.
6. finetune the model.

Before finetune, make sure that the pretrained model is in finetune.sh 's ${output-dir}/checkpoints, and there is a records.jsonl in it to refer to this pretrained model

exp/finetune/checkpoints
├── records.jsonl
└── snapshot_iter_1000000.pdz

The content of records.jsonl should be as follows (change "path" to your own ckpt path):

{"time": "2021-11-21 15:11:20.337311", "path": "~/PaddleSpeech/examples/csmsc/voc3/exp/finetune/checkpoints/snapshot_iter_1000000.pdz", "iteration": 1000000}↩

Run the command below

./finetune.sh

By default, finetune.sh will use conf/finetune.yaml as config, the dump-dir is dump_finetune, the experiment dir is exp/finetune.

TODO: The hyperparameter of finetune.yaml is not good enough, a smaller learning_rate should be used (more milestones should be set).

Pretrained Models

Pretrained model can be downloaded here mb_melgan_baker_ckpt_0.5.zip.

Finetuned model can ben downloaded here mb_melgan_baker_finetune_ckpt_0.5.zip.

Static model can be downloaded here mb_melgan_baker_static_0.5.zip

Model	Step	eval/generator_loss	eval/log_stft_magnitude_loss	eval/spectral_convergence_loss	eval/sub_log_stft_magnitude_loss	eval/sub_spectral_convergence_loss
default	1(gpu) x 1000000	——	——	——	——	——
finetune	1(gpu) x 1000000	3.196967	0.977804	0.778484	0.889576	0.776756

Multi Band MelGAN checkpoint contains files listed below.

mb_melgan_baker_ckpt_0.5
├── default.yaml                  # default config used to train multi band melgan
├── feats_stats.npy               # statistics used to normalize spectrogram when training multi band melgan
└── snapshot_iter_1000000.pdz     # generator parameters of multi band melgan

Acknowledgement

We adapted some code from https://github.com/kan-bayashi/ParallelWaveGAN.