A new paper on arXiv, titled Speech LLMs That Breathe: Modeling Non-Verbal Vocalizations in Dialogue, examines how speech-based large language models handle the non-verbal sounds that punctuate conversation. The abstract notes that user experience has improved thanks to better modeling of paralinguistic attributes and advances in speech reconstruction, but it also points to a gap: human communication follows extensive norms that go beyond words.
The paper appears to argue that while current systems have made progress, they still struggle with the full range of vocal behaviors—such as breaths, laughs, or hesitations—that carry meaning in dialogue. The abstract is truncated, so specific methods and results are not available from the source, but the framing suggests a push toward more natural, norm-aware speech models.
Given the limited information, the key takeaway is that the field recognizes non-verbal vocalizations as a frontier for speech LLMs. The paper likely contributes to ongoing efforts to make spoken interactions with AI feel more human, though the details remain to be seen.