Large language models have been adapted to process speech, creating speech-augmented language models (SLMs) that combine reasoning with spoken input. However, these models can generate hallucinated content, particularly in automatic speech recognition (ASR) tasks, where they may produce transcriptions that are fluent but incorrect.

The new study, available on arXiv, proposes using small language models as a tool to probe and mitigate such hallucinations. The idea is to leverage the smaller models' focused capabilities to identify when the larger SLM's output deviates from the audio, and then correct it.

While the abstract does not provide detailed results, the approach suggests a practical path toward more reliable speech-augmented systems. By pairing a large model with a small verifier, developers may reduce false transcriptions without retraining the entire system.