Stop asking people to pick a language
A meeting in the Gulf routinely runs in Arabic and English inside a single sentence, and the switch is rarely arbitrary. People move to Arabic for nuance, relationships and things that matter, and to English for terms of art, numbers and anything written down.
Forcing one language flattens that. The person who was being precise in Arabic becomes approximate in English, and the meeting gets worse in a way nobody names at the time.
Where transcription tools fall over
Most tools ask you to select a language before recording, then transcribe the whole file as if that choice were true throughout. A meeting that switches produces a transcript that is right for one speaker and garbled for another.
The failure mode worth knowing about: Hindi and Urdu are close enough that a model asked for one will sometimes render the other's script. The words are recognisable, the script is wrong, and the transcript is unusable to the person who needs it. It is the kind of bug that only shows up in a room like ours.
Practical things that improve any recording
Put the microphone nearer the quiet speaker rather than the middle of the table. Rooms have one person everyone strains to hear, and that is the person the transcript will lose.
Say names when people join. "Ahmed, you were going to speak to the contractor" gives a speaker-labelling system something to work with and makes the transcript readable months later.
Say numbers twice in different forms. "Two point four million, so 2.4" survives a bad audio moment in a way a single utterance does not.
Reading a bilingual transcript afterwards
Expect the summary to be in one language even when the transcript is in two. That is usually what you want — a single readable digest — but check that the summary has not quietly dropped the parts it found harder.
The test is simple: find the most important thing that was said in Arabic and check it survived into the summary. If it did not, the tool is transcribing both languages and understanding one.