Conference proceeding
Text Injection for Neural Contextual Biasing
INTERSPEECH 2024, pp.2985-2989
Interspeech
01/01/2024
DOI: 10.21437/Interspeech.2024-1471
Abstract
Neural contextual biasing effectively improves automatic speech recognition (ASR) for crucial phrases within a speaker's context, particularly those that are infrequent in the training data. This work proposes contextual text injection (CTI) to enhance contextual ASR. CTI leverages not only the paired speech-text data, but also a much larger corpus of unpaired text to optimize the ASR model and its neural biasing component. Unpaired text is converted into speech-like representations and used to guide the model's attention towards relevant bias phrases. Moreover, we introduce a contextual text-injected (CTI) minimum word error rate (MWER) training, which minimizes the expected WER caused by contextual biasing when unpaired text is injected into the model. Experiments show that CTI with 100 billion text sentences can achieve up to 43.3% relative WER reduction from a strong neural biasing model. CTI-MWER provides a further relative improvement of 23.5%.
Details
- Title: Subtitle
- Text Injection for Neural Contextual Biasing
- Creators
- Zhong Meng - Google (United States)Zelin Wu - Google (United States)Rohit Prabhavalkar - Google (United States)Cal Peyser - Google (United States)Weiran Wang - Google (United States)Nanxin Chen - Google (United States)Tara N. Sainath - Google (United States)Bhuvana Ramabhadran - Google (United States)
- Resource Type
- Conference proceeding
- Publication Details
- INTERSPEECH 2024, pp.2985-2989
- Series
- Interspeech
- DOI
- 10.21437/Interspeech.2024-1471
- ISSN
- 2308-457X
- Publisher
- Isca-Int Speech Communication Assoc
- Number of pages
- 5
- Language
- English
- Date published
- 01/01/2024
- Academic Unit
- Computer Science
- Record Identifier
- 9984798360202771
Metrics
6 Record Views