Generative Speech Recognition Error Correction with Large Language Models and Task-Activating Prompting

TL;DR

This paper investigates using large language models as post-processors for speech recognition, employing various prompting techniques to improve error correction without extensive fine-tuning, achieving competitive and superior results.

Contribution

It introduces a novel task activation prompting method and demonstrates that LLMs can effectively perform speech recognition rescoring and error correction through prompting alone.

Findings

In-context learning with frozen LLMs achieves competitive rescoring results.

Combining prompting with fine-tuning surpasses N-best oracle error rates.

The proposed methods generalize well across out-of-domain speech recognition tasks.

Abstract

We explore the ability of large language models (LLMs) to act as speech recognition post-processors that perform rescoring and error correction. Our first focus is on instruction prompting to let LLMs perform these task without fine-tuning, for which we evaluate different prompting schemes, both zero- and few-shot in-context learning, and a novel task activation prompting method that combines causal instructions and demonstration to increase its context windows. Next, we show that rescoring only by in-context learning with frozen LLMs achieves results that are competitive with rescoring by domain-tuned LMs, using a pretrained first-pass recognition system and rescoring output on two out-of-domain tasks (ATIS and WSJ). By combining prompting techniques with fine-tuning we achieve error rates below the N-best oracle level, showcasing the generalization power of the LLMs.