Multilingual and Fully Non-Autoregressive ASR with Large Language Model Fusion: A Comprehensive Study

TL;DR

This paper introduces a non-autoregressive, LM-fused ASR system that leverages large language models and parallel processing to significantly reduce latency and improve accuracy across multiple languages.

Contribution

It presents a novel non-autoregressive fusion approach combining USM and PaLM 2, with extensive analysis of parameters affecting performance in large-scale multilingual ASR.

Findings

Achieved 10.8% WER reduction on FLEURS

Achieved 3.6% WER reduction on YouTube captions

Analyzed impact of LLM size, context length, and vocabulary on ASR performance

Abstract

In the era of large models, the autoregressive nature of decoding often results in latency serving as a significant bottleneck. We propose a non-autoregressive LM-fused ASR system that effectively leverages the parallelization capabilities of accelerator hardware. Our approach combines the Universal Speech Model (USM) and the PaLM 2 language model in per-segment scoring mode, achieving an average relative WER improvement across all languages of 10.8% on FLEURS and 3.6% on YouTube captioning. Furthermore, our comprehensive ablation study analyzes key parameters such as LLM size, context length, vocabulary size, fusion methodology. For instance, we explore the impact of LLM size ranging from 128M to 340B parameters on ASR performance. This study provides valuable insights into the factors influencing the effectiveness of practical large-scale LM-fused speech recognition systems.