Evaluation of LLMs in Speech is Often Flawed: Test Set Contamination in Large Language Models for Speech Recognition

TL;DR

This paper reveals that the evaluation of large language models in speech recognition is often flawed due to test set contamination, which biases results and questions the reliability of previous performance claims.

Contribution

It demonstrates the impact of test set contamination on LLM speech evaluation and emphasizes the need for careful data separation in future assessments.

Findings

Contamination in LibriSpeech and Common Voice affects LLM evaluation.

Contaminated LLMs favor training data in their outputs.

Biases in speech recognition results are caused by data contamination.

Abstract

Recent work suggests that large language models (LLMs) can improve performance of speech tasks compared to existing systems. To support their claims, results on LibriSpeech and Common Voice are often quoted. However, this work finds that a substantial amount of the LibriSpeech and Common Voice evaluation sets appear in public LLM pretraining corpora. This calls into question the reliability of findings drawn from these two datasets. To measure contamination impact, LLMs trained with/without contamination are compared. A contaminated LLM is more likely to generate test sentences it has seen during training. Then, speech recognisers based on LLMs are compared. They show only subtle error rate differences if the LLM is contaminated, but assign significantly higher probabilities to transcriptions seen during LLM training. Results show that LLM outputs can be biased by tiny amounts of data contamination, highlighting the importance of evaluating LLM-based speech systems with held-out data.