Detecting different speakers in a session recording: diarization
How speaker diarization separates voices in therapy session recordings—and gives AirNote a clearer, more accurate foundation for AI-generated session notes.
2 min read
A therapy recording may contain several voices, but the audio itself does not arrive with name labels. In an in-person session, therapist and client may share one microphone. A remote session may also be captured as one mixed stream when separate audio sources are not available.
A speech-to-text model can recognise the words accurately while still confusing who said them. If a client’s disclosure is attributed to the therapist—or a therapist’s question is written as the client’s view—the transcript becomes clinically misleading. Any session Note built from it begins on an unstable foundation.
What diarization does
Speaker diarization is the process of working out “who spoke when”. It is separate from transcription: transcription identifies the words, while diarization organises those words into speaker turns.
Dedicated AI models analyse recurring characteristics in the audio—such as patterns in voice and timing—and group segments that appear to come from the same person. The result might initially be labelled Speaker 1 and Speaker 2. Diarization does not need to know a person’s name or understand the clinical meaning of what they said; its job is to keep the voices distinct.
Interruptions and overlapping speech make this a difficult inference rather than a perfect identity test. Even so, reliable separation provides valuable structure for everything that follows.
Why speaker attribution matters
Therapy notes depend not only on what was said but on who said it. A client reporting fear is different from a therapist asking about fear. A client agreeing to a plan is different from a therapist proposing one. The distinction affects how symptoms, interventions, risk, goals and decisions are represented.
When voices are conflated, a summarising model may confidently preserve the wrong relationship between speaker and statement. Better attribution does not replace clinical judgement, but it gives note generation a more faithful account of the session.
How AirNote uses diarization
AirNote preserves speaker information at the earliest reliable point. In remote sessions where therapist and client audio are already available as separate sources, those source boundaries can be retained. When a recording contains a shared or mixed microphone track, AirNote uses a dedicated diarization model on the therapist’s Mac to separate the conversation into speaker turns.
That local processing means the audio does not need to be uploaded to another service simply to identify speakers. AirNote can then use the speaker-aware transcript when generating the Note, helping it distinguish client material from therapist questions, reflections and interventions.
Diarization is not a substitute for the therapist’s review, and no automated separation should be treated as infallible. Its value is quieter and more practical: a clearer transcript, a stronger starting point for the session Note, and less risk that two voices are mistaken for one.
