Publication detail

Reducing Domain mismatch in Self-supervised speech pre-training

BASKAR, M. ROSENBERG, A. RAMABHADRAN, B. ZHANG, Y.

Original Title

Reducing Domain mismatch in Self-supervised speech pre-training

Type

conference paper

Language

English

Original 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.

Keywords

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

Authors

BASKAR, M.; ROSENBERG, A.; RAMABHADRAN, B.; ZHANG, Y.

Released

18. 9. 2022

Publisher

International Speech Communication Association

Location

Incheon

ISBN

1990-9772

Periodical

Proceedings of Interspeech

Number

9

State

French Republic

Pages from

3028

Pages to

3032

Pages count

5

URL

https://www.isca-speech.org/archive/pdfs/interspeech_2022/baskar22_interspeech.pdf

BibTex

@inproceedings{BUT179828,
  author="Murali Karthick {Baskar} and Andrew {Rosenberg} and Bhuvana {Ramabhadran} and Yu {Zhang}",
  title="Reducing Domain mismatch in Self-supervised speech pre-training",
  booktitle="Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH",
  year="2022",
  journal="Proceedings of Interspeech",
  number="9",
  pages="3028--3032",
  publisher="International Speech Communication Association",
  address="Incheon",
  doi="10.21437/Interspeech.2022-736",
  issn="1990-9772",
  url="https://www.isca-speech.org/archive/pdfs/interspeech_2022/baskar22_interspeech.pdf"
}