Healthcare · Clinical AI · Talk!t, Norway

Consultation Summarizer

A clinician talks to a patient. The record writes itself, correctly, in the structure the system expects. The transcription was the easy part. Everything that stops it inventing a symptom is the actual product.

Got something that rhymes with this? · Reply within 24 hours

CONSULTATIONDoctorPatientNothing typedSOAP NOTESOAPChecked, ready
SectorHealthcare
ShapeSpeech plus LLM
HandlingHIPAA and GDPR
Hard partNot the model
Live asTalk!t, Norway
01
The problem

What it replaced.

Clinical notes were written after the fact, from memory, at the end of a shift. The record that resulted was thin, late and occasionally wrong in a way nobody could audit, because there was no artefact to check it against. Every clinician knew this. None of them had time to fix it.

END OF SHIFTWritten from memory, lateNothing to audit it against
The fourth node is why this took months rather than a weekend. A summary that adds a symptom is worse than no summary.
02
What we built

The pieces, and why each one exists.

Transcription that survives real rooms

Two people, overlapping speech, background noise, and a vocabulary where "hypertension" and "hypotension" differ by one phoneme and mean opposite things. General transcription is excellent at conversational English and noticeably worse here, which is where the tuning went.

Structuring into SOAP rather than prose

The output has to land in Subjective, Objective, Assessment and Plan as the record system expects them, not as a paragraph somebody then has to cut up. The model is constrained to that structure rather than asked politely for it.

A verification pass against the transcript

Every clinical claim in the summary is checked back against the words actually spoken. Anything the transcript does not support is dropped rather than softened, and the clinician sees what was dropped.

Handling built in from the first commit

HIPAA and GDPR requirements shaped where audio lives, how long it lives, who can reach it and what is logged. Retro-fitting that after a working prototype is how projects in this sector die.

TRANSCRIPTSUMMARYNot said: droppedClinician sees why
03
The hard part

Where the time actually went.

The failure mode that matters is not a missed word, it is a confident sentence describing something that never happened. A language model asked to summarise will smooth a gap rather than leave one, which is exactly the wrong instinct in a medical record. Most of the engineering is the machinery that makes the system prefer an incomplete note to an invented one, and makes the omission visible to the person signing it.

What we took from it

Carried into later work

  • The clinical safety work is the product, not a phase at the end of it.
  • A model that refuses to guess is more useful here than a model that scores higher.
  • Compliance decided the architecture, so it had to be understood in week one.
The stack

What it runs on.

Speech
Speech to textSpeaker separationClinical vocabulary tuning
Language
LLM pipelineConstrained outputSOAP structuring
Verification
Claim checkingTranscript groundingClinician review step
Compliance
HIPAAGDPRRetention policyAccess logging
04
Live in Norway

The product is Talk!t.

This platform runs in Norway as Talk!t, AI-based clinical documentation for healthcare personnel. Notes are written in real time during the consultation, sorted into the structure a Norwegian journal expects (current issue, findings, assessment, actions), and medical terms are recognised as they are spoken. Nothing has to be installed or learned; it runs in the background.

Patient data is not stored on Talk!t's servers: it is processed in real time and deleted once the note is confirmed in the journal, at the latest after 24 hours, and everything is encrypted. See it live at talkit.no.

Building something like this? See how we work as an AI implementation partner.

JOURNALCurrent issueFindingsAssessmentActionsConfirm noteAudio deleted
Next step

Bring us the problem, not the spec.

Thirty minutes. If your problem rhymes with this one we will say what we would do differently the second time, which is usually the more useful half.