Meetings
Transcribing meetings to Persian text — why it's hard
Transcribing a clean five-minute demo is easy. A real meeting — ninety minutes, two languages, company and people names — breaks somewhere else. These are its four places.
6 min read
The minutes that never get written
Everyone knows something should be written down after the meeting. Nobody writes it — not from laziness, but because the next meeting has already started. So the decisions stay in the attendees' memory and leave with them.
Automatic transcription is the obvious answer, and it always works on the clean five-minute demo. A real Persian meeting has four problems the demo never shows.
Four problems
Where transcribing a real meeting breaks
1. The most valuable words are the hardest ones
People's names, company names and engagement names are what make a transcript usable — and exactly what a general model gets wrong, because they were never in its data. The fix is not a bigger model; it is context: the model knowing who is in this meeting and which company and engagement they are discussing.
2. A Persian meeting is not purely Persian
An English financial term, a Latin product name and a number sit mid-sentence. The transcriber has to know the primary language without "translating" the Latin fragments into Persian — or EBITDA becomes something the reader has to guess.
3. Ninety minutes is not one file
A long meeting cannot be handed to a model whole. It has to be split — on silence boundaries, not mid-word — and each piece transcribed independently, so a failure mid-way resumes from that piece rather than starting over.
4. Models get stuck, and must not get stuck silently
Transcription models have two famous failures: the truncated answer, and the endless repetition of one word. A system that cannot detect both either delivers silently corrupted text or leaves a meeting "processing" forever. Detection, bounded retries and loud failure are part of the architecture, not a detail.
In this product
The meeting is written down as it happens
In PulseUp, the recording arrives at the server in chunks and each chunk is transcribed in the moment — with a glossary built from the meeting itself: the participants, the company, the connected engagement. An already-recorded file is accepted too, split on silence boundaries.
And more important than any accuracy claim: the transcript is a document, under the same rule as every other document — searchable by the people allowed to read it, and only them.
- Live transcription, chunk by chunk — not tomorrow morning
- Each meeting's glossary comes from the meeting itself
- Resumes from the exact chunk that failed
- A model failure is reported, never swallowed
Contract draft review
- The price-adjustment clause reverts to the July formula, with the cap reviewed annually.
- The technical annex arrives from the client next week.
- On the payment schedule, the proposal is…
56 of 120 minutes
Common questions
What is the transcription accuracy?
We don't publish a number — any figure is meaningless without saying what audio it was measured on, and your meeting room sounds nothing like anyone's test studio. Record one real meeting instead and judge the transcript yourself.Does it say who said which sentence?
Not yet. The full text is transcribed, but speaker attribution is being built — and until it ships, we don't claim it.Does it work self-hosted?
Yes. Transcription goes through the same model endpoint you configure, and on a self-hosted deployment it can stay inside your own network.
Try it on your own data
The free plan has no time limit. Bring the hardest document you have.