Weak-Attention Suppression For Transformer Based Speech Recognition

TL;DR

This paper introduces Weak-Attention Suppression (WAS), a dynamic sparsity method for transformers in speech recognition, improving accuracy by reducing attention to redundant frames and emphasizing critical acoustic information.

Contribution

WAS is a novel technique that induces sparsity in attention, leading to state-of-the-art streaming speech recognition performance.

Findings

WAS reduces WER by 10% on test-clean and 5% on test-other.

WAS suppresses attention to non-critical frames, improving model focus.

WAS outperforms strong baselines on LibriSpeech benchmark.

Abstract

Transformers, originally proposed for natural language processing (NLP) tasks, have recently achieved great success in automatic speech recognition (ASR). However, adjacent acoustic units (i.e., frames) are highly correlated, and long-distance dependencies between them are weak, unlike text units. It suggests that ASR will likely benefit from sparse and localized attention. In this paper, we propose Weak-Attention Suppression (WAS), a method that dynamically induces sparsity in attention probabilities. We demonstrate that WAS leads to consistent Word Error Rate (WER) improvement over strong transformer baselines. On the widely used LibriSpeech benchmark, our proposed method reduced WER by 10%$ on test-clean and 5% on test-other for streamable transformers, resulting in a new state-of-the-art among streaming models. Further analysis shows that WAS learns to suppress attention of non-critical and redundant continuous acoustic frames, and is more likely to suppress past frames rather than future ones. It indicates the importance of lookahead in attention-based ASR models.