yanchaomars / pychain

PyTorch implementation of LF-MMI for End-to-end ASR

PyTorch implementation of LF-MMI for End-to-end ASR

End-to-end version of lattice-free MMI (LF-MMI or chain model) implemented in PyTorch.
TODO: regular version of LF-MMI.

What's New:

• GPU computation for both denominator and numerator graphs
• Support unequal length sequences within a minibatch

Installation and Requirements

• PyTorch version >= 1.4.0

First-time Installation (including OpenFST)

git clone https://github.com/YiwenShaoStephen/pychain.git
pip install kaldi_io
make

Update

Whenever you update or modify any none-python codes (e.g. .c or .cu) in pychain, you need to re-compile it by

make pychain

Reference

1. "End-to-end speech recognition using lattice-free MMI", Hossein Hadian, Hossein Sameti, Daniel Povey, Sanjeev Khudanpur, Interspeech 2018 (pdf)
2. "Purely sequence-trained neural networks for ASR based on lattice-free MMI", Daniel Povey, Vijayaditya Peddinti, Daniel Galvez, Pegah Ghahrmani, Vimal Manohar, Xingyu Na, Yiming Wang and Sanjeev Khudanpur, Interspeech 2016, (pdf) (slides,pptx)

The code is based on the original version in kaldi repository with no dependency on the kaldi base.