Audio language models are increasingly considered for deployment in crisis hotlines and customer support, but they only add value over speech-to-text pipelines if they actually use acoustic cues such as tone and prosody. A new arXiv preprint asks whether these models prioritise the transcript when text and paralinguistic signals conflict.

The paper, titled "Text Dominance in Paralinguistic Conflict: Tracing Suppression Through Audio LLM Circuits," describes tests of transcript-versus-acoustic prioritisation and traces how suppression of paralinguistic information might be implemented in the models' internal circuits. The abstract frames this as a direct test of whether audio LLMs go beyond the transcript alone.

The available source is a single preprint abstract, so it does not yet provide detailed results or comparison with other studies. The significance is the question itself: for high-stakes applications like crisis hotlines, ignoring acoustic cues could make audio LLMs behave like ordinary speech-to-text pipelines.