Arabic AI meeting notes — transcription, summary and minutes for Arabic and bilingual meetings
Arabic meeting transcription that keeps the English words
Most meetings in the region are not held in one language. A product name, a tool, a metric or a deadline arrives in English in the middle of an Arabic sentence, and it is usually the word the action item is about. Khulasa transcribes Arabic and English in the same meeting — even inside the same sentence — and keeps those English product names, tools and terms in Latin script instead of rewriting them in Arabic letters.
That is a deliberate choice. A transcript that writes an English word in Arabic letters has not made a small spelling mistake; it has lost the word, and nothing downstream can recover a term that was never written down. The summary, the decisions and the action items are only as good as the transcript underneath them, so the transcript is where the bilingual work is done.
Arabic, English, or both in one sentence
When you request a recording you choose the language once: Auto-detect, Arabic + English, English or Arabic. Auto-detect is the right choice for most meetings; Arabic + English tells the engine to expect both from the first minute; the single-language options are there for meetings you know will stay in one language.
The language is decided once per recording rather than once per turn, so a speaker who switches mid-sentence does not flip the whole transcript into the other language. The summary, decisions and action items then come back in the language that dominated the conversation — an Arabic meeting yields Arabic minutes, an English meeting English minutes, and a mixed one follows whichever side carried most of the talking. We make no claim for specific dialects or for languages other than Arabic and English.
From transcript to minutes: summary, decisions, risks, action items
Every recording produces the same four things. The transcript is the full text of the meeting with speaker labels and timestamps, in Arabic and English exactly as spoken; click any line to jump to that moment in the audio. The summary is a clean, readable account of what was discussed, written in the language that dominated the meeting. Decisions and risks are pulled out as their own lists — what was agreed and what was flagged — so nothing gets lost in the transcript. Action items name who does what, with an owner and a due date where one was mentioned, ready to share or to turn into Jira issues from the meeting page.
Together they are the meeting minutes most teams actually need: not a wall of text to read after the fact, but the outcomes, in the meeting's own language, in your dashboard shortly after the meeting ends, with the summary emailed to whoever requested the recording.
How accurate is it? What we measured
We ran our self-hosted speech engine and every Arabic-focused competitor we could reach through a public API over the same real, bilingual meeting audio — code-switched utterances from UN ESCWA meetings — and scored them all the same way. The whole result is published on the home page, including the columns we do not win.
On overall word error rate this configuration had the lowest figure in the comparison, but two of the four paired confidence intervals span zero, so we report parity with the leading vendors rather than a ranking. On the English words spoken inside Arabic sentences it was more accurate than each of the four competing configurations, one by one, with every interval excluding zero — the one axis where we claim the lead. And it kept over three times as many Arabic–English switch points as any of the three Speechmatics configurations.
Read it as a trade, not a sweep. On the Arabic words alone, all three Speechmatics configurations are ahead of us. We make no per-dialect claim — there are too few clips per dialect in the data to resolve one, and Egyptian Arabic is not covered at all. Every figure describes the self-hosted, on-premise configuration, not our hosted tiers.
An Arabic-first product, not a translation
Khulasa was written for Arabic-speaking teams rather than translated for them afterwards. The app is fully bilingual with a proper right-to-left Arabic interface, so the dashboard, the transcript view and the sharing controls read naturally in either language. When the bot joins a meeting it posts its recording notice in the chat in English and then in Arabic, with a link to the recording notice on khulasa.ai in each language, so an Arabic-speaking participant reads the disclosure in their own words. Both languages are first-class: neither the interface nor the analysis treats Arabic as the second option.
Works in Google Meet, Zoom and Microsoft Teams
The platform does not change what the speech engine hears. Khulasa joins Google Meet from its own Google account and asks the host to admit it; it joins Zoom and Microsoft Teams from the browser as a guest named KhulasaBot (Recording), with nothing installed on the platform side and no Zoom, Microsoft 365 Copilot or Google Workspace licence involved. Paste the meeting link, or connect your Google or Microsoft calendar and let it join on time. Each platform page explains the admit flow, what attendees see and what is not supported.
Google Meet AI notetaker in Arabic and EnglishZoom AI notetaker in Arabic and EnglishMicrosoft Teams AI notetaker in Arabic and English
Keep Arabic meeting audio in the region
Some organisations cannot send meeting audio to a cloud provider at all. For them Khulasa runs as a private cloud — a dedicated, isolated deployment hosted where you need it, the location agreed when we set it up — or on-premise, the same system installed on your own hardware, able to run with no outbound connection. Both replace the managed cloud models with self-hosted speech and language models, so your audio is never sent to a third-party AI provider, and both are quoted per organisation. The benchmark figures above describe exactly that self-hosted configuration.
Questions about Arabic meeting notes
Does the AI notetaker understand Arabic and English in the same meeting?
Yes. Choose Arabic + English, or leave it on Auto-detect, and Khulasa transcribes both languages in one meeting — including sentences that start in Arabic and finish in English — while keeping English names and terms in Latin script. The summary, decisions and action items come back in the language that dominated the conversation.
Is the Arabic transcription as accurate as the English?
Not on every axis, and we publish that. On the Arabic words alone, all three Speechmatics configurations we measured are ahead of our self-hosted engine. Where it leads is on the English words spoken inside Arabic sentences and on keeping the switch points between the two languages, and its overall word error rate came out at parity with the leading vendors. For a bilingual meeting that is the trade we chose; for an Arabic-only broadcast it would not be.
Are the summaries written in natural Arabic?
The summary, decisions and action items are written in the language that dominated the meeting, so an Arabic meeting yields Arabic minutes. English product names, tools and terms stay in Latin script rather than being rewritten in Arabic letters, which is how an Arabic-speaking professional would write them. We do not claim summary quality from a benchmark — word error rate measures the transcript, not the summary.
Which Arabic dialects are supported?
We make no per-dialect claim. The benchmark data has too few clips per dialect to resolve one, Egyptian Arabic is not covered by it at all, and on dialectal Arabic that is not meeting audio the leading vendors are ahead of our configuration. What we measured and publish is bilingual Arabic-and-English meeting audio.
Can I get English notes from an Arabic meeting?
Not as a separate option. The summary follows the recording language you choose when you request the recording: English gives an English summary, Arabic an Arabic one, and Auto-detect or Arabic + English the language that dominated the meeting. There is no separate output-language switch, so if your team needs English notes from an Arabic-language meeting, the transcript — which keeps every English term in Latin script — is the document to work from.
Is there live transcription during the meeting?
No. The bot records the meeting and everything — transcript, summary, decisions, action items — is produced after it ends. Recording starts when the bot is admitted, the audio is uploaded and transcribed once the meeting is over, and the results appear in your dashboard, with the summary emailed to whoever requested the recording.
Can the AI transcribe Arabic meetings accurately?
On the bilingual meeting audio we measured, the self-hosted engine had the lowest overall word error rate in the comparison — at parity with the leading vendors, since two of the four paired intervals span zero — and led on the English words inside Arabic speech. It is behind the Speechmatics configurations on the Arabic words alone. Those are public research corpora, not customer meetings; a real meeting is harder than all of them, and no customer audio has ever been measured.