Publication Details

GPU-Accelerated Forward-Backward Algorithm with Application to Lattice-Free MMI

ONDEL YANG, L.; LAM-YEE-MUI, L.; KOCOUR, M.; CORRO, C.; BURGET, L. GPU-Accelerated Forward-Backward Algorithm with Application to Lattice-Free MMI. In ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings. Singapore: IEEE Signal Processing Society, 2022. p. 8417-8421. ISBN: 978-1-6654-0540-9.
Czech title
Forward-backward algoritmus akcelerovaný na GPU s aplikací pro Lattice-Free MMI trénování
Type
conference paper
Language
English
Authors
Ondel Lucas Antoine Francois, Mgr., Ph.D. (SSDIT)
Lam-Yee-Mui L'ea-Marie
Kocour Martin, Ing. (DCGM)
Corro Caio Filippo
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
URL
Keywords

Lattice-Free MMI, end-to-end ASR, Julia language, forward-backward

Abstract

We propose to express the forward-backward algorithm in terms of operations between sparse matrices in a specific semiring. This new perspective naturally leads to a GPU-friendly algorithm which is easy to implement in Julia or any programming languages with native support of semiring algebra. We use this new implementation to train a TDNN with the LF-MMI objective function and we compare the training time of our system with PyChaina recently introduced C++/CUDA implementation of the LF-MMI loss. Our implementation is about two times faster while not having to use any approximation such as the "leaky-HMM".

Published
2022
Pages
8417–8421
Proceedings
ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings
ISBN
978-1-6654-0540-9
Publisher
IEEE Signal Processing Society
Place
Singapore
DOI
10.1109/ICASSP43922.2022.9746824
UT WoS
000864187908145
EID Scopus
2-s2.0-85130631441
BibTeX 
@inproceedings{BUT178409,
  author="Lucas Antoine Francois {Ondel} and L'ea-Marie {Lam-Yee-Mui} and Martin {Kocour} and Caio Filippo {Corro} and Lukáš {Burget}",
  title="GPU-Accelerated Forward-Backward Algorithm with Application to Lattice-Free MMI",
  booktitle="ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings",
  year="2022",
  pages="8417--8421",
  publisher="IEEE Signal Processing Society",
  address="Singapore",
  doi="10.1109/ICASSP43922.2022.9746824",
  isbn="978-1-6654-0540-9",
  url="https://ieeexplore.ieee.org/document/9746824"
}
Files
Back to top