Feature extraction

This part aims to extract features from speech data, including

• mel-spectrogram
  • ppgvc_mel
  • mel
• linguistic representations
  • vq-wav2vec
  • conformer_ppg
  • hubert_soft and hubert_discrete
• utterance level speaker representations
  • d-vector
• prosodic representations
  • ppgvc_logf0
  • fastspeech2 pitch + energy

Mel-Spectrograms

- ppgvc_mel: logmel_spectrograms from [ppg-vc](https://github.com/liusongxiang/ppg-vc)
- mel: logmel_spectrogram from [parallel_wave_gan](https://github.com/kan-bayashi/ParallelWaveGAN)

ppgvc_mel

It works compatible with the ppgvc_hifigan vocoder. It uses a min max normalization and is trained on VCTK dataset.

./bin/feature_extraction_ppgvc_mel_f0.sh dataset

mel

Linguistic representation extraction

vq-wav2vec