Transformer-based Streaming ASR with Cumulative Attention

TL;DR

This paper introduces a cumulative attention mechanism for streaming Transformer ASR, improving latency and performance by synchronizing attention heads and leveraging accumulated acoustic information.

Contribution

It proposes a novel online attention method called cumulative attention, inspired by MoChA and HS-DACS, with synchronized attention heads to reduce latency in streaming Transformer ASR.

Findings

Achieves comparable or better accuracy than existing streaming Transformer ASR systems.

Reduces inference latency significantly compared to prior methods.

Demonstrates effectiveness on AIShell-1 and Librispeech datasets.

Abstract

In this paper, we propose an online attention mechanism, known as cumulative attention (CA), for streaming Transformer-based automatic speech recognition (ASR). Inspired by monotonic chunkwise attention (MoChA) and head-synchronous decoder-end adaptive computation steps (HS-DACS) algorithms, CA triggers the ASR outputs based on the acoustic information accumulated at each encoding timestep, where the decisions are made using a trainable device, referred to as halting selector. In CA, all the attention heads of the same decoder layer are synchronised to have a unified halting position. This feature effectively alleviates the problem caused by the distinct behaviour of individual heads, which may otherwise give rise to severe latency issues as encountered by MoChA. The ASR experiments conducted on AIShell-1 and Librispeech datasets demonstrate that the proposed CA-based Transformer system can achieve on par or better performance with significant reduction in latency during inference, when compared to other streaming Transformer systems in literature.