April 14, 2024
Large-scale self-supervised pre-trained speech encoders outperform conventional approaches in speech recognition and translation tasks. Due to the high cost of developing these large models, building new encoders for new tasks and deploying them to on-device applications are infeasible. Prior studies propose model compression methods to address this issue, but those works focus on smaller models and less realistic tasks. Thus, we propose Contrastive Layer-to-layer Distillation (CoLLD), a novel knowledge distillation method to compress pre-trained speech encoders by leveraging masked prediction and contrastive learning to train student models to copy the behavior of a large teacher model. CoLLD outperforms prior methods and closes the gap between small and large models on multilingual speech-to-text translation and recognition benchmarks.
Written by
Heng-Jui Chang
Ning Dong (AI)
Ruslan Mavlyutov
Sravya Popuri
Andy Chung
Publisher
ICASSP
May 24, 2024
May 24, 2024
May 06, 2024
Gregoire Mialon , Yann LeCun , Thomas Scialom , Clémentine Fourrier , Thomas Wolf
May 06, 2024
April 22, 2024
Vasu Sharma * , Karthik Padthe * , Newsha Ardalani , Kushal Tirumala , Russ Howes , Hu Xu , Bernie Huang , Daniel Li (FAIR) , Armen Aghajanyan , Gargi Ghosh , Luke Zettlemoyer
April 22, 2024
April 14, 2024
Yun Wang (Speech) , Arthur Hinsvark , Qing He , Shun Zhang , Wonjune Kang
April 14, 2024
Product experiences
Foundational models
Our approach
Research
Product experiences
Latest news
Foundational models