Towards Paradigm-General Suicide Risk Detection via Speech LLM
Conference paper on adolescent suicide-risk detection from speech. Prior speech-based assessment relies on one elicitation paradigm at a time (verbal fluency, reading, question answering); this paper unifies ten elicitation paradigms in a single speech language model using a mixture of DoRA experts that weights complementary cues across assessments. Evaluated on 1,223 participants, the model outperforms paradigm-specific and conventional joint-learning models, generalises to unseen paradigms and shows better confidence calibration.
Publisher
ISCA (Proceedings of Interspeech 2026)
Published
1 Sept 2026
Added
today
Key Findings
- 1,223 participants across ten speech elicitation paradigms, described by the authors as the first cross-paradigm suicide-risk model
- A mixture of DoRA experts on a speech language model beats both paradigm-specific and joint-learning baselines
- Generalises to elicitation paradigms not seen in training
- Improved confidence calibration relative to baselines
Methodology Notes
Five-page Interspeech paper (pages 3852 to 3856, DOI 10.21437/Interspeech.2026-666, not yet registered at Crossref on 2026-09-21). The abstract does not state the age range, the risk-label source (clinical interview or questionnaire), class balance or the size of the unseen-paradigm test; read the PDF before citing any number beyond n=1,223. Author affiliations (Tsinghua University and collaborators) inferred from the author list. Date precision: proceedings posted on the ISCA Archive between 17 and 21 September 2026 ahead of the conference (27 September to 1 October 2026); month-precision convention (day 01) used.
Authors
Li, Jialun, Jiang, Weitao, Cui, Ziyun, Duan, Yinan, Qu, Diyang, Zhang, Chao, Chen, Runsen, Lei, Chang, Wu, Wen
Tags
Cite This
APA
Li, Jialun et al. (2026). Towards Paradigm-General Suicide Risk Detection via Speech LLM. ISCA (Proceedings of Interspeech 2026). https://www.isca-archive.org/interspeech_2026/li26j_interspeech.html
Related Insights
Evaluating Large Language Models in Crisis Detection: A Real-World Benchmark from Psychological Support Hotlines (PsyCrisisBench)
arXiv (Chinese research team) · 2 Jun 2025
Speech-based Psychological Crisis Assessment using LLMs
ISCA (Proceedings of Interspeech 2026) · 1 Sept 2026