Definition
Speaker diarization is the task of automatic audio analysis that divides a recording or audio stream into segments and assigns each segment to a specific speaker. The result is a temporal segmentation of the type "Speaker A: 0:00-0:23, Speaker B: 0:24-1:15, Speaker A: 1:16-2:30." Diarization does not identify who the person is — only that two (or more) distinct people are speaking.
How it's used
In the context of therapy or coaching sessions, diarization is critical for the transcript to be readable and useful. Without it, an hour-long session between therapist and client appears as an undifferentiated block of text. With diarization, the transcript is presented as a structured dialogue: "Therapist: ..., Client: ...", which can be used to generate clinical notes, analyze communication patterns, and detect imbalances in speaking time.
Quality depends on technical factors: audio quality, voice overlaps (when two people speak simultaneously), vocal similarity between speakers, and background noise. In modern video conferences, the separation of audio channels by participant considerably simplifies the task.
In advanced systems, diarization can be combined with speaker identification: once the system "learns" a specific speaker's voice (the therapist), it can consistently label them in future sessions.
When to apply
Diarization is essential in any recording with multiple participants where the authorship of each speech fragment is relevant. It is particularly critical in couples or family therapy sessions (3+ participants), in structured interviews where the candidate's responses need to be distinguished from the interviewer's questions, and in clinical supervision sessions.
Historical origin
Research on speaker diarization began in the 1990s in speech processing laboratories (NIST, DARPA). For decades it was a separate and costly task. Integration of diarization models into commercial transcription systems accelerated significantly in 2018-2022, when neural models reduced diarization error by an order of magnitude.
How CauceOS supports it
CauceOS applies real-time diarization to each session, automatically labeling transcript fragments by speaker. In most cases, the system automatically recognizes whether the audio comes from the professional's or client's channel. The professional can review and correct speaker labels in the post-session transcript view.
Related terms
- Streaming transcription — diarization occurs in parallel with transcription
- Live co-pilot — the co-pilot uses diarization to contextualize its alerts
- Live cross-language translation — diarization is necessary to know which language to translate into in bilingual sessions
References
- Park, T. J., et al. (2022). Review of speaker diarization: Recent advances with deep learning. Computer Speech & Language.
- Fiscus, J. G., et al. (2006). Rich transcription 2006 spring meeting recognition evaluation. Proceedings of RT06.