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

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 the impact of contamination, LLMs trained with or without contamination are compared, showing that a contaminated LLM is more likely to generate test sentences it has seen during training. Speech recognisers using contaminated LLMs shows only subtle differences in error rates, but assigns significantly higher probabilities to transcriptions seen during 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.