Continual learning optimizations for auto-regressive decoder of multilingual ASR systems
Continual Learning (CL) involves fine-tuning pre-trained models with new data while maintaining the performance on the pre-trained data. This is particularly relevant for expanding multilingual ASR (MASR) capabilities. However, existing CL methods, mainly designed for computer vision and reinforc...
| Main Authors: | , , |
|---|---|
| Other Authors: | |
| Format: | Conference Paper |
| Language: | English |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/180315 http://arxiv.org/abs/2407.03645v3 https://www.isca-archive.org/interspeech_2024/index.html |
| _version_ | 1826129204546109440 |
|---|---|
| author | Kwok, Chin Yuen Yip, Jia Qi Chng, Eng Siong |
| author2 | College of Computing and Data Science |
| author_facet | College of Computing and Data Science Kwok, Chin Yuen Yip, Jia Qi Chng, Eng Siong |
| author_sort | Kwok, Chin Yuen |
| collection | NTU |
| description | Continual Learning (CL) involves fine-tuning pre-trained models with new data
while maintaining the performance on the pre-trained data. This is particularly
relevant for expanding multilingual ASR (MASR) capabilities. However, existing
CL methods, mainly designed for computer vision and reinforcement learning
tasks, often yield sub-optimal results when directly applied to MASR. We
hypothesise that this is because CL of the auto-regressive decoder in the MASR
model is difficult. To verify this, we propose four optimizations on the
decoder. They include decoder-layer gradient surgery, freezing unused token
embeddings, suppressing output of newly added tokens, and learning rate
re-scaling. Our experiments on adapting Whisper to 10 unseen languages from the
Common Voice dataset demonstrate that these optimizations reduce the Average
Word Error Rate (AWER) of pretrained languages from 14.2% to 12.4% compared
with Experience Replay, without compromising the AWER of new languages. |
| first_indexed | 2025-03-09T15:02:00Z |
| format | Conference Paper |
| id | ntu-10356/180315 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2025-03-09T15:02:00Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | ntu-10356/1803152024-10-02T00:40:46Z Continual learning optimizations for auto-regressive decoder of multilingual ASR systems Kwok, Chin Yuen Yip, Jia Qi Chng, Eng Siong College of Computing and Data Science Interspeech 2024 Digital Trust Centre Computer and Information Science Continual learning Language-agnostic Continual Learning (CL) involves fine-tuning pre-trained models with new data while maintaining the performance on the pre-trained data. This is particularly relevant for expanding multilingual ASR (MASR) capabilities. However, existing CL methods, mainly designed for computer vision and reinforcement learning tasks, often yield sub-optimal results when directly applied to MASR. We hypothesise that this is because CL of the auto-regressive decoder in the MASR model is difficult. To verify this, we propose four optimizations on the decoder. They include decoder-layer gradient surgery, freezing unused token embeddings, suppressing output of newly added tokens, and learning rate re-scaling. Our experiments on adapting Whisper to 10 unseen languages from the Common Voice dataset demonstrate that these optimizations reduce the Average Word Error Rate (AWER) of pretrained languages from 14.2% to 12.4% compared with Experience Replay, without compromising the AWER of new languages. Info-communications Media Development Authority (IMDA) National Research Foundation (NRF) Submitted/Accepted version This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative 2024-10-02T00:40:46Z 2024-10-02T00:40:46Z 2024 Conference Paper Kwok, C. Y., Yip, J. Q. & Chng, E. S. (2024). Continual learning optimizations for auto-regressive decoder of multilingual ASR systems. Interspeech 2024, 1225-1229. https://dx.doi.org/10.21437/Interspeech.2024-205 2958-1796 https://hdl.handle.net/10356/180315 10.21437/Interspeech.2024-205 http://arxiv.org/abs/2407.03645v3 https://www.isca-archive.org/interspeech_2024/index.html 1225 1229 en © 2024 ISCA (International Speech Communication Association). All rights reserved. This article may be downloaded for personal use only. Any other use requires prior permission of the copyright holder. The Version of Record is available online at http://doi.org/10.21437/Interspeech.2024-205. application/pdf |
| spellingShingle | Computer and Information Science Continual learning Language-agnostic Kwok, Chin Yuen Yip, Jia Qi Chng, Eng Siong Continual learning optimizations for auto-regressive decoder of multilingual ASR systems |
| title | Continual learning optimizations for auto-regressive decoder of multilingual ASR systems |
| title_full | Continual learning optimizations for auto-regressive decoder of multilingual ASR systems |
| title_fullStr | Continual learning optimizations for auto-regressive decoder of multilingual ASR systems |
| title_full_unstemmed | Continual learning optimizations for auto-regressive decoder of multilingual ASR systems |
| title_short | Continual learning optimizations for auto-regressive decoder of multilingual ASR systems |
| title_sort | continual learning optimizations for auto regressive decoder of multilingual asr systems |
| topic | Computer and Information Science Continual learning Language-agnostic |
| url | https://hdl.handle.net/10356/180315 http://arxiv.org/abs/2407.03645v3 https://www.isca-archive.org/interspeech_2024/index.html |
| work_keys_str_mv | AT kwokchinyuen continuallearningoptimizationsforautoregressivedecoderofmultilingualasrsystems AT yipjiaqi continuallearningoptimizationsforautoregressivedecoderofmultilingualasrsystems AT chngengsiong continuallearningoptimizationsforautoregressivedecoderofmultilingualasrsystems |