Building behavioral health documentation that clinicians trust
Clinicians decide whether to trust a behavioral health scribe on the words that carry clinical weight, not your overall accuracy score. Here are the three things that actually earn it.
A therapist finishes a 50-minute session, opens the generated note, and reads that the client "denied SI" — except the client said the opposite, and the model dropped the negation. That note gets deleted. So does the next one. By the third time, the clinician stops opening the tool and goes back to typing at 7pm.
Trust in behavioral health documentation is not won incrementally. It's a threshold. A tool either produces notes a clinician skims and signs, or it produces notes they rewrite from scratch — and the second category gets abandoned regardless of how good the average accuracy number looks in a sales deck.
Behavioral health is also the hardest transcription problem in clinical audio, and it's worth being specific about why. Sessions are long. Speech is emotional, overlapping, and full of trailing sentences. There are frequently more than two people in the room — a couple, a family, a group. Medication names arrive mumbled inside a sentence about something else. And the terminology that matters most is the terminology a general model has seen least.
Here's what it takes to clear the trust threshold, in the order clinicians actually judge it.
Accuracy, measured on the words that carry clinical weight
Word error rate is a bad proxy for behavioral health documentation quality, and leaning on it is how vendors end up with a great benchmark and an abandoned product. A model that transcribes 97% of a session correctly and misses "lamotrigine," "anhedonia," and a negation has failed at the only job that mattered.
The metric that maps to clinical risk is Missed Entity Rate — how often the drugs, conditions, and procedures spoken in the audio never make it into the transcript at all.
With Universal-3.5 Pro and Medical Mode enabled, that number is 3.2%. It's the lowest Missed Entity Rate across the providers we've benchmarked.
Read the methodology and the rest of the set on our benchmarks page.
What the add-on buys you
Compared to the same base model with Medical Mode off, Medical Mode delivers roughly 20% fewer missed medical entities and 87% fewer entity errors. That second number is the one worth internalizing: it's not that the model hears slightly more, it's that the residual error volume in your clinical vocabulary collapses.
Activation is one parameter. There's no separate model to integrate, no second endpoint, no feature regressions to test around.
import assemblyai as aai
aai.settings.api_key = "YOUR_API_KEY"
config = aai.TranscriptionConfig(
speech_models=["universal-3-5-pro"],
domain="medical-v1",
speaker_labels=True,
speakers_expected=2,
)
transcript = aai.Transcriber().transcribe(
"https://example.com/therapy-session.wav",
config=config,
)
The context trick nobody is using
Most teams building behavioral health documentation try to fix vocabulary with word lists — thousands of medication names and diagnostic terms, maintained by hand, uploaded with every request. It works a little and it's a maintenance treadmill.
The better input is already in your database. In an internal healthcare test, feeding a patient's prior-visit note as context cut missed medical terms by 31%. The prior note already names the medications this client takes, the diagnoses in play, and the language this clinician uses. That's a far sharper signal than a generic list of 4,000 drug names, and it's per-patient rather than global.
Speaker separation in sessions that don't take clean turns
A transcript that gets every word right and attributes half of them to the wrong person is unusable for a progress note. Worse, it's dangerous: a note that records the clinician's hypothetical as the client's statement is a documentation error with real consequences.
Behavioral health sessions are unusually hostile to diarization. Clients trail off. Clinicians reflect back what was just said, in similar language, immediately after. In couples and family work, two people share vocal characteristics and interrupt each other constantly. Group sessions can run to eight or more voices.
Why cpWER is the right measure
Diarization Error Rate — the number most vendors publish — scores what fraction of audio time got attributed correctly. It rewards handling long monologues well and barely penalizes dropping short turns. But short turns are where behavioral health content concentrates: "no," "not really," "since Tuesday," "I stopped taking it."
Concatenated minimum-permutation WER scores each speaker's assembled transcript against the truth. It's the measure that reflects what your note-generation step will actually receive. Universal-3.5 Pro's diarization is optimized for cpWER rather than DER, and it's the most accurate diarization we've shipped — specifically stronger on short turns and overlapped speech.
Live sessions versus the note of record
Universal-3.5 Pro Realtime brings diarization to streaming, with revision — the model corrects earlier speaker assignments as the session gives it more evidence — and handles up to 10 speakers. That's what makes group and family work viable live rather than only in post-processing.
Practically, most behavioral health products end up with one of three shapes:
Full rates are on the pricing page. Our take: start post-session. It's cheaper, it's more accurate, and the live view is a feature clinicians ask for less often than product teams assume.
Multilingual sessions, without a language picker
Community behavioral health rarely happens in one language. A client may narrate a childhood memory in Spanish and describe present symptoms in English, sometimes inside the same sentence. Asking the clinician to select a language before the session is both a bad UX and a source of bad transcripts.
The base model code-switches natively across 18 languages with no configuration. Medical Mode's clinical tuning covers English, Spanish, German, and French. Those are two separate things and it pays to keep them separate when you're scoping a deployment: a Spanish-English session gets both; a Vietnamese-English session gets native code-switching without the Medical Mode entity boost.
PHI handling, stated plainly
Behavioral health records carry more sensitivity than most clinical data, and the compliance question usually arrives before the accuracy question in any serious evaluation.
AssemblyAI signs a Business Associate Addendum (BAA) for customers processing PHI, which makes us a business associate under HIPAA for that data. We're SOC 2 Type 2. PHI redaction runs across both audio and transcripts, so what you persist can be stripped of identifiers in both the text and the recording — which matters when your retention policy treats audio differently from text.
The specifics live in the BAA FAQ and the Business Associate Addendum. For teams with data-residency constraints, EU processing and self-hosted deployment are both available.
What's still on you
No API hands you a defensible documentation workflow. The parts that decide whether a compliance review goes well are product decisions: clinician sign-off before a note becomes part of the record, an audit trail on reads and edits, a retention window you can actually justify, and a visible indicator that recording is happening. Teams like Sully AI, Heidi Health, and Deepscribe treat those as first-class product surfaces rather than settings pages, and it shows in adoption.
Where behavioral health documentation goes next
The accuracy problem is close to solved for this use case, which means the interesting work is moving up a layer. The next differentiator isn't a better transcript — it's whether the note sounds like the clinician who signs it. Two therapists documenting the same session write differently, emphasize different things, and use different shorthand. A tool that produces a technically correct note in someone else's voice still gets rewritten.
The teams pulling ahead are building per-clinician style adaptation on top of a transcript they've stopped worrying about, and letting the clinician revise by speaking rather than typing. That's a very different product than a transcription wrapper, and it's the one that clears the trust threshold for good. If you want the implementation view, our walkthrough of how to build an ambient AI scribe for therapy covers the pipeline end to end.
Frequently asked questions
What is the most accurate speech-to-text API for behavioral health transcription?
Universal-3.5 Pro with Medical Mode enabled records a 3.2% Missed Entity Rate on medical entities — the lowest across the providers we've benchmarked. For behavioral health specifically, the other half of the answer is diarization: Universal-3.5 Pro's is optimized for cpWER and handles the short turns and overlapping speech that sessions are full of. More detail on our behavioral health solutions page.
How does Medical Mode improve accuracy on psychiatric medication names?
Medical Mode is tuned for medical entities — drugs, conditions, procedures — and turning it on with domain: "medical-v1" yields roughly 20% fewer missed medical entities and 87% fewer entity errors than the same model without it. For an extra boost, pass the client's prior-visit note as context; in an internal healthcare test that cut missed medical terms by 31%.
How does AssemblyAI handle HIPAA and PHI?
AssemblyAI signs a Business Associate Addendum (BAA) for customers processing PHI, making us a business associate under HIPAA for that data. We're SOC 2 Type 2, and PHI redaction runs across both audio and transcripts. EU data residency and self-hosted deployment are available for teams with stricter requirements.
Can it transcribe couples, family, and group therapy sessions?
Yes. Diarization on pre-recorded audio handles multi-party sessions, and Universal-3.5 Pro Realtime supports up to 10 speakers in streaming with revision, so early speaker assignments get corrected as the session continues. Group work is the case where cpWER-optimized diarization matters most, since so much of the content arrives in short interjections.
What does behavioral health transcription cost per session?
Universal-3.5 Pro is $0.21/hr and Medical Mode adds $0.15/hr, so a 50-minute session transcribed after the fact runs about 30 cents. Streaming is $0.45/hr base, or $0.60/hr with Medical Mode. See the pricing page for the full breakdown.
Does AssemblyAI redact patient identifiers automatically?
Not by default — redaction is opt-in and enabled per request. When you turn it on, PHI redaction covers both the transcript text and the audio, so names, dates, and other identifiers can be removed from the artifacts you store — which lets you keep an unredacted transcript inside a short retention window and persist only the redacted version. The docs cover the available entity types.
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Related posts
