PaddleSpeech/examples/ljspeech/voc0/README.md

WaveFlow with LJSpeech

Dataset

Download the datasaet.

wget https://data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2

Extract the dataset.

tar xjvf LJSpeech-1.1.tar.bz2

Get Started

Assume the path to the dataset is ~/datasets/LJSpeech-1.1. Assume the path to the Tacotron2 generated mels is ../tts0/output/test. Run the command below to

1. source path.
2. preprocess the dataset,
3. train the model.
4. synthesize wavs from mels.

./run.sh

Preprocess the dataset.

./local/preprocess.sh ${preprocess_path}

Train the model

./local/train.sh calls ${BIN_DIR}/train.py.

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

The training script requires 4 command line arguments.

1. --data is the path of the training dataset.
2. --output is the path of the output directory.
3. --ngpu is the number of gpus to use, if ngpu == 0, use cpu.

If you want distributed training, set a larger --ngpu (e.g. 4). Note that distributed training with cpu is not supported yet.

Synthesize

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

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

Synthesize waveform.

1. We assume the --input is a directory containing several mel spectrograms(log magnitude) in .npy format.
2. The output would be saved in --output directory, containing several .wav files, each with the same name as the mel spectrogram does.
3. --checkpoint_path should be the path of the parameter file (.pdparams) to load. Note that the extention name .pdparmas is not included here.
4. --ngpu is the number of gpus to use, if ngpu == 0, use cpu.

Pretrained Model

Pretrained Model with residual channel equals 128 can be downloaded here. waveflow_ljspeech_ckpt_0.3.zip.