Transformer ASR with Contextual Block Processing

TL;DR

This paper introduces a context-aware block processing method for Transformer-based speech recognition, enabling efficient global information modeling and outperforming naive block processing across multiple datasets.

Contribution

It proposes a novel context inheritance mechanism and mask technique for Transformer encoders to improve end-to-end speech recognition performance.

Findings

Outperforms naive block processing on WSJ, Librispeech, VoxForge Italian, and AISHELL-1 datasets.

Effectively encodes global linguistic, channel, and speaker attributes.

Analyzes attention weights to understand global information modeling.

Abstract

The Transformer self-attention network has recently shown promising performance as an alternative to recurrent neural networks (RNNs) in end-to-end (E2E) automatic speech recognition (ASR) systems. However, the Transformer has a drawback in that the entire input sequence is required to compute self-attention. In this paper, we propose a new block processing method for the Transformer encoder by introducing a context-aware inheritance mechanism. An additional context embedding vector handed over from the previously processed block helps to encode not only local acoustic information but also global linguistic, channel, and speaker attributes. We introduce a novel mask technique to implement the context inheritance to train the model efficiently. Evaluations of the Wall Street Journal (WSJ), Librispeech, VoxForge Italian, and AISHELL-1 Mandarin speech recognition datasets show that our proposed contextual block processing method outperforms naive block processing consistently. Furthermore, the attention weight tendency of each layer is analyzed to clarify how the added contextual inheritance mechanism models the global information.