FFTNet

A PyTorch implementation of the FFTNet: a Real-Time Speaker-Dependent Neural Vocoder

Quick Start:

1. Install requirements:

pip install -r requirements.txt

2. Download dataset:

wget http://festvox.org/cmu_arctic/cmu_arctic/packed/cmu_us_slt_arctic-0.95-release.tar.bz2
tar xf cmu_us_slt_arctic-0.95-release.tar.bz2

3. Extract features:

python preprocess.py

4. Training with default hyperparams:

python train.py

5. Synthesize from model:

   python generate.py --checkpoint=/path/to/model --lc_file=/path/to/local_conditon

TODO:

• Test and modify conditonal sampling.

• Fast generation. (Without fast generation, it generates about 230 samples per second with a free TITAN Xp.)

• Post-synthesis denoising.

Notes:

• This is not offical implementation, some details are different from the paper.

• Now the generated speech from this repo is not so good as wavenet vocoder.

• Work in progress.