How do you use Mistral Voxtral 3 to transcribe a meeting in another language — and when does a voice agent beat typing?

Mistral’s Voxtral 3 pairs speech-to-text in 50+ languages (with speaker diarization and context biasing for names and jargon) with text-to-speech in 35+ languages, priced from $0.0035 a minute. Here is how to transcribe a real meeting with it, what the terms mean in plain English, and the one check that keeps transcripts trustworthy.

A supplier call in French, a standup in Spanish, and notes to write afterwards in English — that juggling act is exactly what voice AI is now priced to end. Short answer: Voxtral 3 is Mistral’s new voice toolkit — speech-to-text that Mistral says covers more than 50 languages, labels who is speaking, and can be primed with your names and jargon in advance, plus text-to-speech in 35+ languages for the reply. Record the meeting, upload the audio, add your context list, and you get a readable transcript in the meeting’s own language, ready to summarize or translate.

Two terms worth knowing, in plain English. Diarization means the transcript tags each speaker — “Speaker 1,” “Speaker 2” — instead of one undifferentiated wall of text, which is what makes a four-person call readable. Context biasing means you hand the system a list of proper nouns and terms — people’s names, your company, product names — before it listens, so it spells them your way instead of guessing.

For live use, Mistral says its realtime model answers in under half a second, and its voice-agent tooling can tell when you have finished a sentence.1. Get consent, then record. Tell everyone on the call it is being recorded.

A clean recording helps any model more than clever settings do.2. Pick your route. Use Mistral’s own studio or API, or an app built on Voxtral — developers also get a voice-agent SDK built on LiveKit for live calls.3.

Upload and set the language — or let it detect among its 50+ languages automatically.4. Add your context list before you run it. Names, acronyms, product terms.

This is the step everyone skips and everyone regrets.5. Read it with the speaker labels. Rename “Speaker 1” to the actual person once, and the whole transcript falls into place.6.

Know the cost. Mistral lists Mini Transcribe 3 at $0.0035 a minute — about 21 cents for an hour-long meeting — and its realtime model at $0.008 a minute. Transcription has quietly become one of the cheapest things AI does.When does a voice agent beat typing?

When the conversation itself is the work — calls, interviews, multilingual meetings, or hands-busy moments — and when you need the answer spoken back in another language, which the same pipeline handles.Pro tip: Always spot-check names, numbers and dates against the audio. Diarization tells you who spoke, not that every word is right — and a confidently wrong figure in a quote or deadline is the expensive kind of error.Ad’s take: The headline is not the price, it is the boring parts being handled — who spoke, and how the names spell. That is what turns a transcript from a data dump into something you can actually use.