Publication Details

Language Model Integration Based on Memory Control for Sequence to Sequence Speech Recognition

CHO Jaejin, WATANABE Shinji, HORI Takaaki, BASKAR Murali K., INAGUMA Hirofumi, VILLALBA Lopez Jesus Antonio and DEHAK Najim. Language Model Integration Based on Memory Control for Sequence to Sequence Speech Recognition. In: Proceedings of 2019 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP). Brighton: IEEE Signal Processing Society, 2019, pp. 6191-6195. ISBN 978-1-5386-4658-8. Available from: https://ieeexplore.ieee.org/document/8683380
Czech title
Integrace jazykového modelu v sequence-to-sequence rozpoznávání řeči založená na řízení paměťových buněk
Type
conference paper
Language
english
Authors
Cho Jaejin (JHU)
Watanabe Shinji, Dr. (JHU)
Hori Takaaki (MERL)
Baskar Murali K. (DCGM FIT BUT)
Inaguma Hirofumi (KyotoUni)
Villalba Lopez Jesus Antonio (JHU)
Dehak Najim (JHU)
URL
Keywords

Automatic speech recognition (ASR), sequence to sequence, language model, shallow fusion, deep fusion, cold fusion

Abstract

In this paper, we explore several new schemes to train a seq2seq model to integrate a pre-trained language model (LM). Our proposed fusion methods focus on the memory cell state and the hidden state in the seq2seq decoder long short-term memory (LSTM), and the memory cell state is updated by the LM unlike the prior studies. This means the memory retained by the main seq2seq would be adjusted by the external LM. These fusion methods have several variants depending on the architecture of this memory cell update and the use of memory cell and hidden states which directly affects the final label inference. We performed the experiments to show the effectiveness of the proposed methods in a mono-lingual ASR setup on the Librispeech corpus and in a transfer learning setup from a multilingual ASR (MLASR) base model to a low-resourced language. In Librispeech, our best model improved WER by 3.7%, 2.4% for test clean, test other relatively to the shallow fusion baseline, with multilevel decoding. In transfer learning from an MLASR base model to the IARPA Babel Swahili model, the best scheme improved the transferred model on eval set by 9.9%, 9.8% in CER, WER relatively to the 2-stage transfer baseline.

Published
2019
Pages
6191-6195
Proceedings
Proceedings of 2019 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP)
Conference
2019 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), Brighton, GB
ISBN
978-1-5386-4658-8
Publisher
IEEE Signal Processing Society
Place
Brighton, GB
DOI
UT WoS
000482554006084
EID Scopus
BibTeX
@INPROCEEDINGS{FITPUB12252,
   author = "Jaejin Cho and Shinji Watanabe and Takaaki Hori and K. Murali Baskar and Hirofumi Inaguma and Antonio Jesus Lopez Villalba and Najim Dehak",
   title = "Language Model Integration Based on Memory Control for Sequence to Sequence Speech Recognition",
   pages = "6191--6195",
   booktitle = "Proceedings of 2019 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP)",
   year = 2019,
   location = "Brighton, GB",
   publisher = "IEEE Signal Processing Society",
   ISBN = "978-1-5386-4658-8",
   doi = "10.1109/ICASSP.2019.8683380",
   language = "english",
   url = "https://www.fit.vut.cz/research/publication/12252"
}
Back to top