Towards Effective and Efficient Non-autoregressive Decoding Using Block-based Attention Mask

TL;DR

This paper introduces a block-based attention mask decoder for Conformer ASR systems that improves decoding speed without sacrificing accuracy by balancing parallel non-autoregressive inference and autoregressive predictions.

Contribution

It proposes a novel AMD module with a dynamic beam search that effectively combines NAR and AR decoding for improved efficiency and performance.

Findings

Achieves up to 1.73x speed-up over baseline decoding.

Maintains statistically insignificant WER increase.

Reduces WER by up to 0.7% absolute with same real-time factors.

Abstract

This paper proposes a novel non-autoregressive (NAR) block-based Attention Mask Decoder (AMD) that flexibly balances performance-efficiency trade-offs for Conformer ASR systems. AMD performs parallel NAR inference within contiguous blocks of output labels that are concealed using attention masks, while conducting left-to-right AR prediction and history context amalgamation between blocks. A beam search algorithm is designed to leverage a dynamic fusion of CTC, AR Decoder, and AMD probabilities. Experiments on the LibriSpeech-100hr corpus suggest the tripartite Decoder incorporating the AMD module produces a maximum decoding speed-up ratio of 1.73x over the baseline CTC+AR decoding, while incurring no statistically significant word error rate (WER) increase on the test sets. When operating with the same decoding real time factors, statistically significant WER reductions of up to 0.7% and 0.3% absolute (5.3% and 6.1% relative) were obtained over the CTC+AR baseline.