Publication Details

Reducing Domain mismatch in Self-supervised speech pre-training

BASKAR Murali K., ROSENBERG Andrew, RAMABHADRAN Bhuvana and ZHANG Yu. Reducing Domain mismatch in Self-supervised speech pre-training. In: Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH. Incheon: International Speech Communication Association, 2022, pp. 3028-3032. ISSN 1990-9772. Available from: https://www.isca-speech.org/archive/pdfs/interspeech_2022/baskar22_interspeech.pdf
Type
conference paper
Language
english
Authors
Baskar Murali K. (DCGM FIT BUT)
Rosenberg Andrew (Google, Inc.)
Ramabhadran Bhuvana (Google, Inc.)
Zhang Yu (Google, Inc.)
URL
Keywords

Self-supervision, Wav2vec2, pretraining, Data selection, Domain mismatch, asr, speech recognition

Abstract

Masked speech modeling (MSM) methods such as wav2vec2 or w2v-BERT learn representations over speech frames which are randomly masked within an utterance. While these methods improve performance of Automatic Speech Recognition (ASR) systems, they have one major limitation. They treat all unsupervised speech samples with equal weight, which hinders learning as not all samples have relevant information to learn meaningful representations. In this work, we address this limitation. We propose ask2mask (ATM), a novel approach to focus on specific samples during MSM pre-training. ATM employs an external ASR model or scorer to weight unsupervised input samples by performing a fine-grained data selection. ATM performs masking over the highly confident input frames as chosen by the scorer. This allows the model to learn meaningful representations. We conduct fine-tuning experiments on two well-benchmarked corpora: LibriSpeech (matching the pre-training data) and, AMI and CHiME-6 (not matching the pre-training data). The results substantiate the efficacy of ATM on significantly improving the recognition performance under mismatched conditions while still yielding modest improvements under matched conditions.

Published
2022
Pages
3028-3032
Journal
Proceedings of Interspeech - on-line, no. 9, ISSN 1990-9772
Proceedings
Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH
Conference
Interspeech 2022, Incheon Korea, KR
Publisher
International Speech Communication Association
Place
Incheon, KR
DOI
EID Scopus
BibTeX
@INPROCEEDINGS{FITPUB12853,
   author = "K. Murali Baskar and Andrew Rosenberg and Bhuvana Ramabhadran and Yu Zhang",
   title = "Reducing Domain mismatch in Self-supervised speech pre-training",
   pages = "3028--3032",
   booktitle = "Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH",
   journal = "Proceedings of Interspeech - on-line",
   number = 9,
   year = 2022,
   location = "Incheon, KR",
   publisher = "International Speech Communication Association",
   ISSN = "1990-9772",
   doi = "10.21437/Interspeech.2022-736",
   language = "english",
   url = "https://www.fit.vut.cz/research/publication/12853"
}
Back to top