Transcribing therapy sessions: what speech recognition does and does not do
A transcription converts speech into text. It recognises words; it does not understand them. The result is a raw text containing everything that was said — filler words, self-corrections and recognition errors. Those who know the difference between this raw text and professional documentation can use both sensibly; those who overlook it carry errors over.
Why the topic matters
Transcription is usually talked about as if it were solved. In psychotherapy practice it is not quite: two people speak in turn and sometimes at the same time, technical terms and medication names come up, and people speak quietly. What the system makes of this determines how much follow-up work remains — which is why it pays to understand the technology rather than trust it.
How the process works
Recognising, not understanding
A speech recognition system assigns sound sequences to words, supported by a language model for the language concerned. It reflects what was said — not what was meant. No assessment of content takes place in this step.
Separating speakers
Speaker separation assigns contributions to different voices. It works on acoustic features, not on roles: the system knows that two people are speaking, not which of them is the one providing treatment.
Raw text, not clean text
The result contains filler words, broken-off sentences, self-corrections — and, where recognition was uncertain, plausible-sounding wrong words. The latter is the real pitfall: a recognition error does not look like an error.
Basis, not result
The raw text is material. What of it belongs in the treatment documentation is a professional decision, not a processing step.
What AI can do here
- Convert spoken language into text, even across an entire session.
- Separate contributions by different speakers.
- Recognise technical vocabulary, insofar as the language model used covers it.
- Provide the raw text as the basis for later condensing.
What AI does not do here
It does not understand what is being talked about.
Source: General property of speech recognition — not a product statement
It does not recognise without errors, and its errors do not look like errors.
Source: General property of speech recognition — not a product statement
It does not know which speaker is the one providing treatment.
Source: Speaker separation works on acoustic features, not on roles
A transcript is not professional documentation.
Source: Knowledge: AI-assisted documentation in psychotherapy
How Psymira implements the process
Psymira transcribes via a cloud service from Google, including speaker separation; processing takes place in the EU. Which service and which model exactly are used is stated in the privacy policy. The transcript remains editable and can be deleted separately from the recording — the steps that follow are described in the help section.
Data protection and professional control
The required prior consent of everyone recorded must be in place for the recording; that is a requirement for the practice, not a setting in the software. Audio files are deleted after successful transcription, at the latest after 30 days. Before AI calls, identifying master data are reduced; pseudonymous codes are used instead of real names. Free text such as notes and transcripts may still contain names — control over this lies with the practice.
Full details on sub-processors, processing locations and retention periods on Privacy & security.
Sources
Primary sources for the legal and technical statements on this page. Product statements are marked above with their origin in our own material.
- Google Cloud Speech-to-Text — documentation
The service used, including language models and speaker separation.
- § 630f BGB (German Civil Code) — documentation of treatment (in German)
What treatment documentation is — and therefore what a transcript is not.
- Regulation (EU) 2016/679 (GDPR), Art. 9 and Art. 28
Health data as a special category and processing on behalf.
- BPtK — practice information “Administrative KI in Ihrer Praxis” (February 2026, in German)
Guidance from the German Federal Chamber of Psychotherapists on the responsible use of administrative AI systems — data protection and professional law requirements, EU AI Act, Medical Device Regulation. It explicitly points out that important questions in the AI context have not yet been settled in a legally binding way.