Dual Causal/Non-Causal Self-Attention for Streaming End-to-End Speech Recognition

TL;DR

This paper introduces the dual causal/non-causal self-attention (DCN) architecture for streaming end-to-end speech recognition, improving performance and enabling frame-synchronous processing compared to previous self-attention methods.

Contribution

The paper proposes the DCN architecture that balances causal and non-causal self-attention, enhancing streaming ASR performance and enabling frame-synchronous processing.

Findings

DCN outperforms restricted self-attention in streaming ASR tasks.

The proposed method achieves state-of-the-art results on LibriSpeech, HKUST, and Switchboard.

DCN provides a good trade-off between context and latency in streaming ASR.

Abstract

Attention-based end-to-end automatic speech recognition (ASR) systems have recently demonstrated state-of-the-art results for numerous tasks. However, the application of self-attention and attention-based encoder-decoder models remains challenging for streaming ASR, where each word must be recognized shortly after it was spoken. In this work, we present the dual causal/non-causal self-attention (DCN) architecture, which in contrast to restricted self-attention prevents the overall context to grow beyond the look-ahead of a single layer when used in a deep architecture. DCN is compared to chunk-based and restricted self-attention using streaming transformer and conformer architectures, showing improved ASR performance over restricted self-attention and competitive ASR results compared to chunk-based self-attention, while providing the advantage of frame-synchronous processing. Combined with triggered attention, the proposed streaming end-to-end ASR systems obtained state-of-the-art results on the LibriSpeech, HKUST, and Switchboard ASR tasks.