✨ Supervoice GPT

February 26, 2024 · View on GitHub

A GPT model that converts from text to phonemes with durations that is suitable to feed into voice synthesizer.

How it works?

This model converts raw text to phonemes and their durations compatible with Montreal Forced Aligner.

This model converts string like "Hey, Vera, what time is it?" to list of tuples of phoneme and it's duration: [('ç', 9), ('iː', 7), ('v', 7), ('ɛ', 8), ('ɹ', 8), ('i', 7), ('w', 6), ('ɐ', 5), ('ʔ', 3), ('tʰ', 8), ('aj', 11), ('m', 7), ('ɪ', 6), ('z', 7), ('ɪ', 6), ('ʔ', 8)]

Dataset

This module require extensive dataset preparation. To prepare all needed data next commands are required to be performed:

  1. datasets sync to download datasets
  2. python ./datasets_prepare.py to preprocess audio files and extract texts from datasets
  3. ./datasets_align.sh to generate alignments
  4. python ./datasets_mix.py to mix all data together
  5. python ./train_tokenizer.py to train tokenizer on alignments
  6. python ./datasets_tokenize.py to tokenize datasets

Training

To train network execute:

./train.sh

License

MIT