Arabic and English in the same meeting

A meeting with a Lebanese client rarely happens in one language. The sentence starts in Arabic, the technical term is in English, the deadline is back in Arabic. TrackHeros transcribes those meetings on Google Meet and Microsoft Teams as they were actually spoken, keeps the real speaker names, and writes the notes in the language you choose for each client. This page is for anyone whose client calls switch between Arabic and English, and for anyone who has read a caption transcript of one and recognised nothing.

Why does single-language transcription break?

Live captions decode one language at a time. Tell them the meeting is in English and every Arabic phrase is forced into the nearest English sounds: nonsense words, confidently punctuated. Tell them Arabic and the English terms get the same treatment. A meeting that switches inside a sentence — which is how bilingual people actually talk — is the worst case, because the captions cannot switch with it. The transcript goes wrong precisely where the substance is: the number, the name, the decision.

Most meeting tools inherit this from the captions they rely on, then summarise the result, and the summary reads as fluently as if it were true.

What does it do?

• Transcribes Arabic and English as spoken, including switches mid-sentence.
• Keeps the real speaker names from the meeting platform.
• Records which languages were detected in each transcript and flags mid-sentence switching.
• Writes the transcript and the notes in the language you set per client (default English).
• Listens for the languages you expect per client (default English and Arabic).
• Remembers the language of a recurring series, so an English-only weekly call does not pay for work it does not need.
• Shows Arabic as Arabic script, read right-to-left, next to the English — not transliterated into Latin letters.

How does it work?

Two passes over every meeting.

Pass	Source	What it contributes
Captions	The platform's live captions, collected by the bot during the call	Real speaker names and timing. Decoded in one language, so they are the scaffold, not the record
Audio	The recording itself, transcribed after the call	The authoritative transcript: both languages, in the order they were spoken, aligned to the captions' speakers

The captions come first and cheaply, and they are what gives the transcript its names, because the platform knows who is speaking. The audio pass is what actually reads a bilingual sentence. The final transcript is the audio pass's text with the captions' speakers attached. A line that cannot be attributed to a speaker is labelled Unknown rather than assigned to whoever seems likeliest.

Correction rules help here too. The people in the room and the client's own terms are told to the transcriber in advance, so a product called UBA does not keep coming out as UBVA — see correction rules.

What is recorded about the languages?

Each transcript records which languages were detected and whether speakers switched mid-sentence. Those flags sit on the transcript itself, so months later you can see that a meeting was bilingual and that the audio pass ran. They also feed the series memory described below.

Which language are the notes written in?

Two settings, both per client, because a developer with a Beirut client and a London client needs two answers and a global setting can only give one.

Setting	Default	What it does
Expected languages	English and Arabic	What the transcription listens for
Output language	English	What the transcript and the notes are written in

A client in London with an Arabic-speaking team gets English notes. A client who reads Arabic gets Arabic notes: set that client's output language to Arabic and the summary, decisions and action items are written in Arabic. The output language is always included in the expected list, because a meeting held in Arabic still contains "staging" and "deploy", and a transcriber told to expect only Arabic would try to write those as Arabic words. The minutes template is then applied in that language.

How do recurring meetings learn their language?

The audio pass is far more expensive than captions, and an English-only meeting gains nothing from it but the cost. A per-client switch would be the wrong tool: one client can hold a bilingual daily standup and an English-only weekly review, and a single flag is wrong for one of them whichever way it is set.

So language is remembered per recurring series. The same people, meeting about the same thing, speak the same way each week. Once a series has been heard several times with only the output language spoken, its later occurrences skip the audio pass and keep the captions. One sighting of a second language — a single Arabic sentence in an otherwise English call — turns the audio pass back on for that series at once, whatever the count of clean occurrences before it, and it stays on rather than being argued down by the next quiet week.

Every doubt resolves toward transcribing. A series that has never been seen takes the full pass. A series whose memory cannot be read takes the full pass. Being wrong toward audio costs money, which is recoverable; being wrong toward captions costs the meeting, which is not. The memory is per client as well as per series, so "Daily standup" under two clients is two separate memories that never inform each other.

How is Arabic shown in the transcript?

As Arabic script, reading right-to-left within the line, next to the English that surrounds it — so a bilingual transcript reads the way it was spoken rather than as Arabic sounds spelled out in Latin letters. The guide to transcribing Arabic and English meetings has a worked example, and the freelance developer page covers how this fits the rest of a bilingual practice.

What does it not do yet?

More languages are on the roadmap; English and Arabic are what TrackHeros promises today.