The same board meeting, run through both. Where we match, where we lose, and what closes each gap.
Two things are genuinely different. Plaud hands the whole meeting back in English, translating the Arabic; ours writes Arabic as it was spoken and now also produces a full English version of the same meeting, so you can read it either way. And Plaud writes a summary, which ours does not do yet - that is the one place they are clearly ahead. Neither transcript is the truth, so everything below compares the two against each other, not against what was really said.
| What matters | Plaud | Ours | Call | What closes the gap |
|---|---|---|---|---|
| How much of the meeting survives | 9,466 words | 9,692 words as spoken, or 11,593 in the English version | ours | Nothing needed. Checked five minutes at a time across the whole two hours: there is no stretch one version heard and the other missed. |
| Arabic | Translated into English. Under 2% of their file is Arabic script. | Arabic kept exactly as spoken, plus a separate all-English version of the same meeting. | your choice | Already built and already run on this meeting. The English version costs about a third more of the Mac's time: 21 minutes per hour of meeting instead of 15. |
| Names | The reference. 37 names come up more than once in their file. | 32 of those 37 are in our English version. Of the five that are not, two are the same person spelled our way (Muhammad for Mohammed, Yazeed for Yazid) and two are not names at all. | even | Give the engine a short list of the people and companies who come up in your meetings. It reads that list before each run, which is the standard way to pin spellings like Sirar, STC and Al Juhfani. In the Arabic version the names are all there too, written in Arabic script. |
| Numbers | Figures throughout. | 95% of Plaud's figures appear in ours at the same moment in the English version, and 100% in the English stretches of the as-spoken version. | even | Nothing pressing. Arabic speech written as Arabic words rather than digits is why the as-spoken version scores lower; the English version already turns them back into figures. |
| Telling voices apart | Twelve speakers. | Twelve voices, agreeing with Plaud's labelling 88.7% of the time. | even | Tell it how many people were in the room and it sharpens further. Putting real names on the voices, learned once and remembered, is the next piece of work. |
| Written summary and action list | A structured summary with sections, conclusions and participants. | None yet. | Plaud wins | Two ways, your call. Either your Claude writes it from the transcript text when you ask, which would be the only thing ever leaving the Mac and never the audio. Or a text model runs on the Mac itself, about 5 GB, an ordinary download with no account, and writes it with nothing leaving at all. |
| What it costs and where it goes | Their cloud, a subscription, and your board audio uploaded to it. | About 15 minutes of your Mac's time per hour of meeting, 21 with the English version. No fee, no account, nothing uploaded, tested with every network call refused. | ours | Nothing needed. |
Trimmed on purpose. The full text of your meeting stays on the Mac.
| Version | When | What changed |
|---|---|---|
| 1 | 2 Aug | First side-by-side on the 4 Aug board recording |