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

Training and inference scripts for FCL-taco2


  • 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

Training and inference:

  • Step1. Data preparation & preprocessing
  1. Download LJSpeech

  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. Or you can download our alignment results, 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:

  2. Training student model FCL-taco2-S:

  3. Parallel-WaveGAN vocoder training: follow instructions at here. 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:

  2. FCL-taco2-S evaluation: