It was shown recently that a combination of ASR and TTS models yield highly competitive performance on standard voice conversion tasks such as the Voice Conversion Challenge 2020 (VCC2020). To obtain good performance both models require pretraining on large amounts of data, thereby obtaining large models that are potentially inefficient in use. In this work we present a model that is significantly smaller and thereby faster in processing while obtaining equivalent performance. To achieve this the proposed model, Dynamic-GAN-VC (DYGAN-VC), uses a non-autoregressive structure and makes use of vector quantised embeddings obtained from a VQWav2vec model. Furthermore dynamic convolution is introduced to improve speech content modeling while requiring a small number of parameters. Objective and subjective evaluation was performed using the VCC2020 task, yielding MOS scores of up to 3.86, and character error rates as low as 4.3%. This was achieved with approximately half the number of model parameters, and up to 8 times faster decoding speed.
- clone repository
git clone https://github.com/MingjieChen/DYGANVC.git
cd DYGANVC- Create conda env and install pytorch 1.7 through conda
conda create --name torch_1.7 python==3.7
conda activate torch_1.7
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
conda install librosa -c conda-forgeChoose your own cudatoolkit version according to your own GPU.
- install packages
pip install fairseq parallel_wavegan munch pyyaml SoundFile tqdm scikit-learn tensorboardX webrtcvad- download dataset and unzip dataset to vcc20/
./data_download.sh vcc20- extract speaker embeddings
./extract_speaker_embed.sh- download vqwav2vec ckpt
mkdir vqw2v
cd vqw2v
wget https://dl.fbaipublicfiles.com/fairseq/wav2vec/vq-wav2vec_kmeans.pt- extract vqwav2vec features
./vqw2v_feat_extract.sh- start training
./run_train.sh- inference
python inference.py