Conference proceeding
Improving Rare Word Recognition with LM-aware MWER Training
INTERSPEECH 2022, Vol.2022-, pp.1031-1035
Interspeech
01/01/2022
DOI: 10.21437/Interspeech.2022-10660
Abstract
Language models (LMs) significantly improve the recognition accuracy of end-to-end (E2E) models on words rarely seen during training, when used in either the shallow fusion or the rescoring setups. In this work, we introduce LMs in the learning of hybrid autoregressive transducer (HAT) models in the discriminative training framework, to mitigate the training versus inference gap regarding the use of LMs. For the shallow fusion setup, we use LMs during both hypotheses generation and loss computation, and the LM-aware MWER-trained model achieves 10% relative improvement over the model trained with standard MWER on voice search test sets containing rare words. For the rescoring setup, we learn a small neural module to generate per-token fusion weights in a data-dependent manner. This model achieves the same rescoring WER as regular MWER-trained model, but without the need for sweeping fusion weights.
Details
- Title: Subtitle
- Improving Rare Word Recognition with LM-aware MWER Training
- Creators
- Weiran Wang - Google Inc, Mountain View, CA 94043 USATongzhou Chen - GoogleTara N. Sainath - Google Inc, Mountain View, CA 94043 USAEhsan Variani - GoogleRohit Prabhavalkar - GoogleRonny Huang - Google Inc, Mountain View, CA 94043 USABhuvana Ramabhadran - GoogleNeeraj Gaur - Google Inc, Mountain View, CA 94043 USASepand Mavandadi - Google Inc, Mountain View, CA 94043 USACal Peyser - Google Inc, Mountain View, CA 94043 USATrevor Strohman - GoogleYanzhang He - GoogleDavid Rybach - Google
- Resource Type
- Conference proceeding
- Publication Details
- INTERSPEECH 2022, Vol.2022-, pp.1031-1035
- Publisher
- Isca-Int Speech Communication Assoc
- Series
- Interspeech
- DOI
- 10.21437/Interspeech.2022-10660
- ISSN
- 2308-457X
- eISSN
- 1990-9772
- Number of pages
- 5
- Language
- English
- Date published
- 01/01/2022
- Academic Unit
- Computer Science
- Record Identifier
- 9984696561702771
Metrics
1 Record Views