arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2604.21928cs.CL

使用生成式大语言模型评估自动语音识别

Evaluation of Automatic Speech Recognition Using Generative Large Language Models

Idiap Research Institute(Idiap研究 institute);EPFL(瑞士联邦理工学院);Brno University of Technology(布拉格技术大学);Avignon University(阿维尼翁大学);Le Mans University(勒曼大学);Nantes University(南特大学)

Thibault Bañeras-Roux, Shashi Kumar, Driss Khalil, Sergio Burdisso, Petr Motlicek, Shiran Liu, Mickael Rouvier, Jane Wottawa, Richard Dufour

AI总结:

本文提出利用生成式大语言模型通过假设选择、语义距离计算和错误分类三种方法评估ASR,在HATS数据集上达到92-94%的人类一致性,优于WER和语义指标。

摘要:

自动语音识别(ASR)传统上使用词错误率(WER)进行评估,该指标对语义不敏感。基于嵌入的语义指标与人类感知的相关性更好,但基于解码器的大语言模型(LLM)在此任务中仍未得到充分探索。本文通过三种方法评估其相关性:(1)在两个候选假设中选择最佳假设,(2)使用生成式嵌入计算语义距离,以及(3)对错误进行定性分类。在HATS数据集上,最佳LLM在假设选择中与人类标注者的一致性达到92-94%,而WER仅为63%,并且也优于语义指标。来自基于解码器的LLM的嵌入显示出与编码器模型相当的性能。最后,LLM为可解释且语义化的ASR评估提供了有前景的方向。

英文摘要:

Automatic Speech Recognition (ASR) is traditionally evaluated using Word Error Rate (WER), a metric that is insensitive to meaning. Embedding-based semantic metrics are better correlated with human perception, but decoder-based Large Language Models (LLMs) remain underexplored for this task. This paper evaluates their relevance through three approaches: (1) selecting the best hypothesis between two candidates, (2) computing semantic distance using generative embeddings, and (3) qualitative classification of errors. On the HATS dataset, the best LLMs achieve 92--94\% agreement with human annotators for hypothesis selection, compared to 63\% for WER, also outperforming semantic metrics. Embeddings from decoder-based LLMs show performance comparable to encoder models. Finally, LLMs offer a promising direction for interpretable and semantic ASR evaluation.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑