Probing and Mitigating Hallucinations in Speech-augmented Language Models for Automatic Speech Recognition via Small Language Models

The widespread success of large language models in natural language processing has spurred the development of speech-augmented language models (SLMs) to extend their reasoning capabilities from textual inputs to spoken ones. Despite impressive performance, SLMs remain susceptible to hallucinations, posing challenges in practical utility. To this end, we first probe hallucinations in SLMs for automatic speech recognition (ASR) via the lens of causal mediation and behavioral analysis. To mitigate fabricated ASR hypotheses, we propose AudioSLM, a compact yet effective ASR framework built upon a small language model, which leverages alignment cues and cross-attention layers to synergistically bolster acoustic modeling and cross-modal interactions. Experiments on LibriSpeech reveal that hallucinations arise from multi-head self-attention modules with an excessive attention bias towards textual tokens. AudioSLM considerably eases ASR hallucinations and outperforms some LLM-based ASR models.