Mistral used Voxtral 3 context biasing to caption AI Engineer Paris

Mistral used Voxtral 3 Transcribe to caption every main-stage talk at AI Engineer Paris in real time. The company shared the demo on X on November 9, 2026. To get speaker names and jargon right, it fed the model a list of names and technical terms from the conference agenda. The feature behind that trick is called context biasing, and Mistral says the same setup is open to anyone building events or apps.

The post is a product demo, not a launch. Voxtral 3 came out on October 20, and we covered its benchmarks and prices at the time. What the post adds is a look at how Mistral runs it in a live setting, and at the one feature that matters most there.

What Mistral showed from AI Engineer Paris

AI Engineer Paris took place on September 23 and 24, 2026, at Station F. Mistral organized the conference with support from the AI Engineer team. That makes the captions a home test: Mistral ran its own model at its own event, about four weeks before Voxtral 3 was public.

The 43-second video in the post is not footage from the stage. It shows a demo app with two panels. On the left sits a list of 100 drug names, from acetazolamide to beclomethasone, one term per line. Each line is sent to the model as one context_bias entry. On the right, the live transcript of a mock call to a doctor appears as the caller speaks.

The model writes "beclomethasone" and "triamcinolone" correctly as they are spoken. It also handles a name spelled out letter by letter. Labels mark where each turn starts and ends, and a counter shows 7 turns with an average hangover of 413 ms. That is the short wait before the model decides a speaker has finished.

The conference and the doctor demo share one problem: words a general model rarely hears.

How context biasing works in Voxtral 3

Context biasing lets a developer hand the model a list of words before it starts listening. In the Mistral API it is the context_bias parameter, and it works on all three Voxtral 3 transcription models. Each entry can be a key phrase, such as a speaker's full name, or a short description of the recording. Both kinds can go in one list.

The docs put the limit plainly. Context changes "which words the model prefers". It does not change "what the model hears". If a term is not in the audio, the list will not make it appear. If a speaker mumbles, the list only nudges the model toward the expected spelling.

How the list is sent depends on the mode:

In realtime mode the list cannot change once audio starts, and the docs recommend key phrases over descriptions there. Mistral does not set a fixed maximum number of terms for Voxtral 3. The list shares a 65,536-token window with the audio, and a request that runs over fails with an error. The cap of 100 terms belonged to the previous generation, Voxtral Transcribe 2.

Mistral's own numbers show what the feature buys. On earnings calls, recall of key terms rises from 81.4% without context to 94.6% with it. No third party has repeated that test yet.

For a conference, that list is easy to build. The agenda already holds every speaker name, company and product that will come up on stage.

What to know before using it for an event

Voxtral Mini Transcribe Realtime 3 costs $0.008 per minute. A two-day event with eight hours of main-stage talks would cost under $4 to caption, by our math. The model is in public preview, and Mistral has not released its weights.

Latency is a setting, not a fixed number. The docs let a session pick a delay between 80 ms and 2,400 ms. Mistral's charts show fewer errors at longer delays, so live captions on a screen and a voice agent on a phone call need different choices.

A few limits still apply:

• Speaker labels and turn detection cannot run in the same session
• Realtime segments carry speaker labels but no start or end times
• Batch output has timestamps per segment, not per word, unlike Voxtral Transcribe 2

Rival models offer similar keyword lists, so the feature itself is not unique. Mistral's case rests on accuracy. Voxtral Small Transcribe 3 ranks third on the Artificial Analysis speech-to-text leaderboard, at a 2.1% word error rate.

Is Voxtral 3 ready for live event captions?

For a conference full of names and jargon, the setup Mistral describes is the right one, and it is cheap. The term list does the work a human captioner would do with a prep sheet. The open question is how well it held up on stage. Mistral has not published the AI Engineer Paris transcripts or an error rate for them, and the video shows a staged doctor's call instead. Until those transcripts are out, the demo proves the workflow and leaves the live accuracy to Mistral's word.

Sources

• Voxtral 3 Transcribe at AI Engineer Paris, Mistral AI on X – November 9, 2026
• Voxtral 3 announcement, Mistral – October 20, 2026
• Bias context and vocabulary, Mistral docs – how context_bias works
• Audio overview, Mistral docs – latency, timestamps and realtime limits
• AI Engineer Paris 2026 – dates and venue
• Speech-to-text leaderboard, Artificial Analysis – AA-WER, read November 9, 2026