torch-nansy

Torch implementation of NANSY, Neural Analysis and Synthesis, arXiv:2110.14513

• Neural Analysis and Synthesis: Reconstructing Speech from Self-Supervised Representations, Choi et al., 2021. [arXiv:2110.14513]

Requirements

Tested in python 3.7.9 conda environment.

Usage

Initialize the submodule and patch.

git submodule init --update
cd hifi-gan; patch -p0 < ../hifi-gan-diff

Download LibriTTS[openslr:60], LibriSpeech[openslr:12] and VCTK[official] datasets.

Dump the dataset for training.

python -m speechset.utils.dump \
    --out-dir ./datasets/dumped

To train model, run train.py

python train.py \
    --data-dir ./datasets/dumped

To start to train from previous checkpoint, --load-epoch is available.

python train.py \
    --data-dir ./datasets/dumped \
    --load-epoch 20 \
    --config ./ckpt/t1.json

Checkpoint will be written on TrainConfig.ckpt, tensorboard summary on TrainConfig.log.

tensorboard --logdir ./log

To inference model, run inference.py

python inference.py \
    --ckpt ./ckpt/libri100_73.ckpt \
    --hifi-ckpt ./ckpt/hifigan/g_02500000 \
    --hifi-config ./ckpt/hifigan/config.json \
    --context ./sample1.wav \
    --identity ./sample2.wav

[TODO] Pretrained checkpoints will be relased on releases.

To use pretrained model, download files and unzip it. Followings are sample script.

from nansy import Nansy

ckpt = torch.load('t1_200.ckpt', map_location='cpu')
nansy = Nansy.load(ckpt)
nansy.eval()

Learning curve and Figures

Epoch: 65

[TODO] Samples