Speech-based Psychological Crisis Assessment using LLMs

Key Findings

• — Three-way crisis-level classification from real hotline speech: macro-F1 0.802, accuracy 0.805 under five-fold cross-validation
• — Acoustic cues rendered as text evidence let a text-only language model use vocal affect that is lost in speech-recognition transcripts
• — Model-generated diagnostic rationales used as auxiliary supervision regularise the classifier
• — Outperforms acoustic-only, zero-shot language-model and speech-aware language-model baselines on the same data
• — Framed around operator variability in hotline crisis assessment under training and staffing constraints

Methodology Notes

Five-page Interspeech paper (pages 5506 to 5510, DOI 10.21437/Interspeech.2026-997, not yet registered at Crossref on 2026-09-21). The abstract does not state the number of calls, the hotline or the label source; the author list includes Yongsheng Tong (Beijing Suicide Research and Prevention Center) and Chao Zhang (Tsinghua University), so the data are probably the Beijing hotline corpus behind the held PsyCrisisBench row; confirm affiliations and sample size from the PDF before citing numbers. Single dataset, cross-validation rather than a held-out deployment test. Date precision: the ISCA Archive posted the Interspeech 2026 proceedings between 17 and 21 September 2026, ahead of the conference (27 September to 1 October 2026, Sydney); the paper page carries no day, so the month-precision convention (day 01) is used.

Sources