Distilling Knowledge from Ensembles of Acoustic Models for Joint CTC-Attention End-to-End Speech Recognition

TL;DR

This paper enhances end-to-end speech recognition by extending multi-teacher knowledge distillation with novel strategies that incorporate error rate metrics, achieving state-of-the-art results across multiple datasets and languages.

Contribution

It introduces three new distillation strategies that integrate error rate metrics into multi-teacher distillation for joint CTC-attention ASR systems, improving recognition performance.

Findings

Achieved state-of-the-art error rates on Common Voice French, Italian, and TIMIT datasets.

Demonstrated effectiveness of error-based teacher selection strategies across multiple languages.

Validated the proposed methods on diverse datasets and training procedures.

Abstract

Knowledge distillation has been widely used to compress existing deep learning models while preserving the performance on a wide range of applications. In the specific context of Automatic Speech Recognition (ASR), distillation from ensembles of acoustic models has recently shown promising results in increasing recognition performance. In this paper, we propose an extension of multi-teacher distillation methods to joint CTC-attention end-to-end ASR systems. We also introduce three novel distillation strategies. The core intuition behind them is to integrate the error rate metric to the teacher selection rather than solely focusing on the observed losses. In this way, we directly distill and optimize the student toward the relevant metric for speech recognition. We evaluate these strategies under a selection of training procedures on different datasets (TIMIT, Librispeech, Common Voice) and various languages (English, French, Italian). In particular, state-of-the-art error rates are reported on the Common Voice French, Italian and TIMIT datasets.