charlottecuc / semi-tts

Semi-supervised Learning for Multi-speaker Text-to-speech Synthesis Using Discrete Speech Representation

Semi-supervised TTS

Semi-supervised Learning for Multi-speaker Text-to-speech Synthesis Using Discrete Speech Representation

Data

Prepare data

Download VCTK and LJSpeech and put it into data/audio-corpus. Specifically, waves from LJSpeech should be in data/audio-corpus/lj and waves from speaker p225 of VCTK should be in data/audio-corpus/p225, etc.. The data partition is specified in data/partition_tables/<partition-table.csv>. The phoneme transcription of each wave file is in data/map_tables/lj_vctk_g2p.csv.

For members in NTU speech lab, the audio-corpus could be downloaded from /groups/public/ttao/audio-corpus.zip.

Audio preprocessing

The hyperparameter for audio features could be modified in config/<config.yaml>. The audio preprocessing code is in src/audio.py.

Model

To adjust the model hyperparameters or learning rate, please modify the configure file in config/<config.yaml>.

Running

Train from scratch

python main.py --config config/<config.yaml> --njobs <num-workers>

Continue training

python main.py --config config/<config.yaml>\
               --njobs <num-workers>\
               --load <checkpoint-path>

The training log could be found in directory log/.

For members in NTU speech lab, the checkpoints could be downloaded from /groups/public/ttao/semi-tts-ckpt.zip.

Inference

python main.py --gen-specgram\
               --config config/<config.yaml>\
               --njobs <num-workers>\
               --load <checkpoint-path>\
               --logdir <output-directory>

Reference

• Grapheme-to-phoneme tool: https://github.com/Kyubyong/g2p
• Tacotron-2: https://github.com/NVIDIA/tacotron2
• End-to-end ASR: https://github.com/Alexander-H-Liu/End-to-end-ASR-Pytorch

Citation

@inproceedings{liu2020towards,
  title={Towards unsupervised speech recognition and synthesis with quantized speech representation learning},
  author={Liu, Alexander H and Tu, Tao and Lee, Hung-yi and Lee, Lin-shan},
  booktitle={ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={7259--7263},
  year={2020},
  organization={IEEE}
}

@article{tu2020semi,
  title={Semi-supervised Learning for Multi-speaker Text-to-speech Synthesis Using Discrete Speech Representation},
  author={Tu, Tao and Chen, Yuan-Jui and Liu, Alexander H and Lee, Hung-yi},
  journal={arXiv preprint arXiv:2005.08024},
  year={2020}
}