an implement of wavenet vocoder using tensorflow

!!! the audio code is copied from wavenet_vocoder !!!

!!! the main tensorflow model is fixed from tensorflow-wavenet !!!

Some issue

mixture is in the branch of dev, but there are some bugs in generating wavs.

To Do

• local condition
• global condition
• multi speaker
• multi gpu
• use conv1d transposed to upsample
• mixture logistic distribution (doing..)
• Tacotron + Wavenet Vocoder

Required

• python >= 3.3
• tensorflow > =1.3
• tqdm
• pyworld
• pysptk
• nnmnkwii >= 0.12
• scipy
• lws == 1.0

Getting Start

Download dataset

• the voice conversion dataset(for multi speaker, 16k): cmu_arctic
• the single speaker dataset(22.05k): LJSpeech-1.0

Preprocess data

for train faster, we should process the data to npy

python preprocess.py --num_workers 4 --name ljspeech --in_dir /your_path/LJSpeech-1.0 --out_dir /your_outpath/ --hparams sample_rate=22050

Training

for single speaker

python train.py --num_gpus 4 --batch_size 2 --train_txt /your_train_txt/ --hparams gc_enable=False,global_channel=0,global_cardinality=0,NPY_DATAROOT=/your_npy_datadir/,sample_rate=22050 --logdir_root log_ljspeech

for multi speaker

python train.py --batch_size 2 --num_gpus 4 --train_txt /your_train_txt/ --logdir_root log_arctic

Synthesize

for single speaker

the eval_txt is extracted from the train_txt

python mul_generate.py --eval_txt /your_eval_txt/ --wav_out_path test_ljspeech.wav /your_cheakpoint/ ---hparams gc_enable=False,global_channel=0,global_cardinality=0,NPY_DATAROOT=/your_npy_datadir/,sample_rate=22050

for multi speaker

python mul_generate.py --eval_txt /your_eval_txt/ --wav_out_path test_arctic.wav /your_checkpoint/ --gc_id 6