Block diagram of FCL-taco2, where the decoder generates mel-spectrograms in AR mode within each phoneme and is shared for all phonemes.

• python 3.6.10
• torch 1.3.1
• chainer 6.0.0
• espnet 8.0.0
• apex 0.1
• numpy 1.19.1
• kaldiio 2.15.1
• librosa 0.8.0

• Step1. Data preparation & preprocessing

1. Download LJSpeech from https://keithito.com/LJ-Speech-Dataset/

2. Unpack downloaded LJSpeech-1.1.tar.bz2 to /xx/LJSpeech-1.1

3. Obtain the forced alignment information by using Montreal forced aligner tool https://montreal-forced-aligner.readthedocs.io/en/latest/. Or you can download our alignment results at https://drive.google.com/file/d/1nMjPlPDtspPxgG5AK-AdBq1ce330YMGP, then unpack it to /xx/TextGrid

4. Preprocess the dataset to extract mel-spectrograms, phoneme duration, pitch, energy and phoneme sequence by:

    python preprocessing.py --data-root /xx/LJSpeech-1.1 --textgrid-root /xx/TextGrid
   

• Step2. Model training

1. Training teacher model FCL-taco2-T:

    ./teacher_model_training.sh
   

2. Training student model FCL-taco2-S:

    ./student_model_training.sh
   

3. Parallel-WaveGAN vocoder training: follow instructions at https://github.com/kan-bayashi/ParallelWaveGAN. You can also download the pre-trained PWG vocoder, and put the PWG model under the directory "vocoder".

• Step3. Model evaluation

1. FCL-taco2-T evaluation:

    ./inference_teacher.sh
   

2. FCL-taco2-S evaluation:

    ./inference_student.sh
   

If you find this repository helpful for you, please cite our paper:

@inproceedings{wang2021fcl,
  title={Fcl-Taco2: Towards Fast, Controllable and Lightweight Text-to-Speech Synthesis},
  author={Wang, Disong and Deng, Liqun and Zhang, Yang and Zheng, Nianzu and Yeung, Yu Ting and Chen, Xiao and Liu, Xunying and Meng, Helen},
  booktitle={ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={5714--5718},
  year={2021},
  organization={IEEE}
}