Collaborative Training of Acoustic Encoders for Speech Recognition

TL;DR

This paper introduces a collaborative training method for acoustic encoders of varying sizes in speech recognition, leveraging shared knowledge and co-distillation to improve performance across models.

Contribution

It proposes a joint training approach with shared modules and co-distillation for different-sized encoders, enhancing efficiency and accuracy in speech recognition models.

Findings

Up to 11% relative WER improvement on LibriSpeech

Shared encoder training benefits multiple model sizes

Co-distillation enhances frame-level prediction accuracy

Abstract

On-device speech recognition requires training models of different sizes for deploying on devices with various computational budgets. When building such different models, we can benefit from training them jointly to take advantage of the knowledge shared between them. Joint training is also efficient since it reduces the redundancy in the training procedure's data handling operations. We propose a method for collaboratively training acoustic encoders of different sizes for speech recognition. We use a sequence transducer setup where different acoustic encoders share a common predictor and joiner modules. The acoustic encoders are also trained using co-distillation through an auxiliary task for frame level chenone prediction, along with the transducer loss. We perform experiments using the LibriSpeech corpus and demonstrate that the collaboratively trained acoustic encoders can provide up to a 11% relative improvement in the word error rate on both the test partitions.