Improving multilingual speech recognition systems

Gao, Heting

Permalink

https://hdl.handle.net/2142/113890 Copy

Description

Title

Improving multilingual speech recognition systems

Author(s)

Gao, Heting

Issue Date

2021-12-03

Director of Research (if dissertation) or Advisor (if thesis)

Hasegawa-Johnson, Mark

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• multilingual speech recognition
• fairness
• language embedding
• zero-shot,

Abstract

End-to-end trainable deep neural networks have become the state-of-the-art architecture for automatic speech recognition (ASR), provided that the network is trained with a sufficiently large dataset. However, many existing languages are too sparsely resourced for deep learning networks to achieve as high accuracy as their resource-abundant counterparts. Multilingual recognition systems mitigate data sparsity issues by training models on data from multiple language resources to learn a speech-to-text or speech-to-phone model universal to all languages. The resulting multilingual ASR models usually have better recognition accuracy than the models trained on the individual dataset. In this work, we propose that two limitations exist for multilingual systems, and resolving the two limitations could result in improved recognition accuracy: (1) existing corpora are of the considerably varied form (spontaneous or read speech), corpus size, noise level, and phoneme distribution and the ASR models trained on the joint multilingual dataset have large performance disparities over different languages. We present an optimizable loss function, equal accuracy ratio (EAR), that measures the sequence-level performance disparity between different user groups and we show that explicitly optimizing this objective reduces the performance gap and improves the multilingual recognition accuracy. (2) While having good accuracy on the seen training language, the multilingual systems do not generalize well to unseen testing languages, which we refer to as cross-lingual recognition accuracy. We introduce language embedding using external linguistic typologies and show that such embedding can significantly increase both multilingual and cross-lingual accuracy. We illustrate the effectiveness of the proposed methods with experiments on multilingual and multi-user and multi-dialect corpora.

Graduation Semester

2021-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/113890

Copyright and License Information

Copyright 2021 Heting Gao

Owning Collections