Abstract
Suicide risk assessment is an inevitable yet inherently difficult task to carry out partly due to the reliance of traditional methodology on the subjective estimation that is usually lengthy, time-consuming, and labor-intensive. This study introduces a multimodal suicide risk assessment automated segmentation pipeline specifically designed based on a set of 103 structured clinical video interview corpus (including 69 control participants and 34 at-risk participants). All of them were automatically divided into 1,834 utterance-level units using auditory extraction, automated speech recognition based on Whisper, and speaker diarization based on PyAnnote. The pipeline provides time-congruent transcripts with speaker labels, thus enabling a more accomplished examination of respondent behavior. The speaker diarization addition was vital in isolating the speech of the subject accurately and eliminating the possibility of interviewer biasing. On the whole, this method is scalable, effective in its operations, and ecologically sound, thus, providing a reproducible framework that can support the future developments of multimodal suicide risk detection and contribute to the general discussion of using speech signal processing in mental health studies.