Skip to content

Latest commit

 

History

History
30 lines (25 loc) · 1.2 KB

README.md

File metadata and controls

30 lines (25 loc) · 1.2 KB
Raw

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:

  • August 2020: GPU computation for graphs in log domain (recommended for numerator graphs)
  • April 2020: Support unequal length sequences within a minibatch
  • April 2020: Examples of using PyChain: Espresso and pychain-example
  • January 2020: GPU computation for both denominator and numerator graphs

Installation and Requirements

First-time Installation (including OpenFST)

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

"PyChain: A Fully Parallelized PyTorch Implementation of LF-MMI for End-to-End ASR", Yiwen Shao, Yiming Wang, Daniel Povey and Sanjeev Khudanpur (pdf)