G2P model

Training

Example: to train en_us arpabet model.

• Place cmudict-0.7b file in en_us folder.
• Run commands:

cd py
pip install -r requirements.txt
python g2p/train.py

To train a different model:

• Create a new folder next to en_us. Place your dictionary in it.
• Copies cfg.yaml into it and modifies graphemes and phonemes to match your language. The first 4 graphemes and phonemes must be <unk>, <pad>, <bos>, <eos>.
• Modify train.py to load your config and dictionary.
• You may need to reformat your dictionary so that SphinxDataset can load it. Or you can write your own dataset class.
• You will need to tweak batch size and epochs (and maybe other parameters) for best results.

Packing

A g2p pack zip file contains:

dict.txt
g2p.onnx
phones.txt

• The G2pPack class uses entries from dict.txt first.
• If not found, it uses g2p.onnx to generate phonemes.
• phones.txt allows phonemizers to know which phonemes are vowels and which phonemes to stretch.