The meeting ends with a summary and tasks, not a recording
A live Hebrew transcript, who-said-what attribution, and key points you can click to jump straight to the moment.
The short answer
How do you transcribe a Hebrew meeting and get a summary and action items automatically?
You start recording from the browser at the beginning of the meeting, and a live transcript builds as it runs; when it ends the system works out who spoke (diarization), replaces the live transcript with a speaker-separated one, and writes a summary, key points and action items — each carrying a timestamp you can click to jump to that exact moment in the recording. With Daber AI this is built Hebrew-first, you can replace "Speaker 1" with a real name, and meetings live in the same system as phone calls — meaning the same PII masking, the same role-based access and the same audit log. Organisations that must keep material inside can run transcription and diarization on their own infrastructure.
From a recording to a transcript you can actually read
During the meeting a live transcript builds with timestamps, so anyone joining late can see what has happened so far. The heavy pass happens after it ends: the system waits for the complete recording, runs speaker diarization, and replaces the live transcript with a speaker-separated one — text you can read like minutes rather than a stream of words. That split is deliberate: the live pass optimises for speed, the second pass for accuracy and structure.
- A live transcript during the meeting, with timestamps
- A second pass that replaces it with a speaker-separated transcript
- The pass can be re-run on an existing meeting
- The recording is stored alongside the transcript, access controlled by role
A summary, key points and clickable action items
On top of the transcript, the system writes a short summary, key points and a list of action items. What makes that useful is that every item points back at the moment it came from: click an action item and the recording jumps exactly there. This addresses the real problem with meeting notes — not "what was said" but "did we actually agree to this, and who said it". If the recording was started without a title, one is written automatically, without ever overwriting a title someone typed.
Speakers with names, not numbers
Diarization separates speakers and labels them "Speaker 1", "Speaker 2" and so on — it works out that these are different people, not who they are. After the meeting you can give each label a real name, and the name applies across the whole transcript. The mapping is stored separately from the transcript, so labels stay stable even if the pass is re-run, and a wrongly entered name can be cleared. That is also the honest statement of the limit: the system does not identify a person from their voice, it separates speakers.
Privacy: a meeting is a record like any other
Because meetings sit in the same model as calls, every privacy mechanism applies to them with no separate configuration: identifying details are detected and masked in the transcript, access to the recording is role-controlled, and every administrative action is written to the audit log. A deletion request runs through the same anonymisation flow — the transcript is genuinely redacted, the recording is detached, and a reference number is kept to quote back to the person who asked. An organisation running the system itself can choose local models for transcription and diarization, so audio never leaves the perimeter.
What this is not, and where it struggles
It records from the browser — what is said in the room or in the session you opened — rather than being a bot that joins Zoom or Teams on your behalf and sits there as a participant. Transcript quality depends on the microphone: a meeting captured by one speakerphone in a large room, or three people talking at once, will hurt both transcription and speaker separation — something to fix with hardware, not software. And finally, this is not a certified legal record: the transcript and summary are an excellent working tool, but if you need binding minutes, a person has to review and approve them.
Frequently asked
Can transcription run on our own servers?
Yes — that is the default posture for regulated organisations. Transcription and diarization run on local models inside your perimeter, so audio is never sent to an external provider. You can also choose an external provider where no such constraint exists and you would rather not run the infrastructure, and that decision is made per department rather than for the whole organisation at once.
Why does the live transcript differ from the final one?
Because they are produced by two deliberately different passes. The live transcript optimises for speed and is shown while the meeting runs; the second pass gets the complete recording, works out who spoke, and produces a more accurate and structured transcript that replaces it. If you re-run the pass, the speaker names you assigned are preserved.
Does it handle a meeting that mixes in English?
Yes. The foundation is Hebrew-first, including Israeli terminology and accents, and English phrases or terms inside a Hebrew conversation are handled as part of the transcript. What actually hurts accuracy is audio quality — background noise, speakers far from the microphone, or people talking over each other — far more than language mixing.
Who can view and listen to a meeting after it ends?
Access is by role, and listening to the recording is more restricted than reading the transcript — playback for ordinary viewers is off by default and can be enabled for a department that wants it. Every administrative view and export action is written to the audit log, so "who accessed this" is a question you can answer without guessing.
Call the agent now
A real call, in Hebrew. Exactly what your callers would hear.
Have the agent call me
Leave a number — the agent calls you in under a minute.
Updated: