Voxtral, Mistral AI’s speech recognition: how good is it for professional dictation?

Mistral AI, the French company founded in Paris in 2023, made its name with its language models. Since July 2025, it has also been developing Voxtral, a family of speech recognition models. TechVoice uses it for all its transcriptions, both live and for uploaded files. Here is what you need to know about Voxtral, and what we learned by measuring it on professional dictation in French.

Voxtral at a glance

• Developer: Mistral AI, Paris.
• Modes: transcription of audio files (batch) and real-time transcription (streaming).
• Access: through Mistral’s API, hosted in Europe; the first generations are also available as open weights.
• Pricing: per minute of audio, in the region of a few tenths of a US cent.

Three generations in a year and a half

When Voxtral 3 came out, Mistral’s list prices were $0.0035 per minute for Mini Transcribe 3, $0.004 for Small Transcribe 3 and $0.008 for real-time transcription. An hour of recording transcribed by the Small model therefore costs about 24 US cents.

The thirteen languages of Voxtral Transcribe 2 were English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian and Dutch. Voxtral 3 extends coverage to more than fifty languages for file transcription, including the twenty-four official languages of the European Union, and to slightly fewer in real time.

File or real time: two ways of listening

File transcription receives the whole recording. The model has all the context before it writes anything, which makes it more accurate. This is the right mode for a voice memo or a recorded meeting.

Real-time transcription receives the audio in small chunks as you speak, and sends the text back as it goes. One setting, the target delay, trades responsiveness against accuracy: the longer the model waits before writing, the better it hears. In our measurements, going from 1,000 to 500 milliseconds cost 7 points of error on specialist terms, without making the transcription finish any sooner. We stayed at 1,000 ms.

TechVoice uses Voxtral Mini Transcribe Realtime 3 for live dictation and Voxtral Small Transcribe 3 for uploaded files, with the language set in both cases to that of the app: French or English.

What we measured

Our test bench replays orthopaedic reports dictated by a synthetic voice and counts, on top of the overall error rate, the specialist terms missed. That is the metric that decides, because a report in which dyspnée (dyspnoea) turns into dispersée (dispersed) is unusable, even if everything else is perfect.

These figures are optimistic: a synthetic voice articulates perfectly and knows neither the background noise of a consulting room nor hesitation. They are for comparing options with one another, not for predicting the error rate on your own voice. The jump in quality between generations, however, is clear and consistent from one measurement to the next.

One technical detail shows the kind of surprise you only find by measuring. Opening a real-time session takes about a second on Mistral’s side, and the model does not catch up on the audio that builds up in the meantime: the first word arrived at around 3.5 seconds. By keeping a session open in advance, ready for the next dictation, TechVoice displays it at around 2.5 seconds, with no change in quality.

Limits to be aware of

1. Context biasing is limited. Voxtral accepts a list of 100 terms to favour, with no spaces within a term: “heart failure” has to be written with an underscore, which the model sometimes copies into its output. Mistral describes this mechanism as experimental outside English; on our French dictations, it added nothing after correction and made live dictation worse. We do not use it.
2. One hour per file. The documentation states a maximum of 60 minutes. In practice, the audio and the text produced share a window of 65,536 tokens: a 104-minute file was rejected. TechVoice caps uploads at 60 minutes.
3. Real-time sessions can drop. A session can be closed at any moment on the provider’s side (“service restart”): one of our probes stopped after 7 min 33 s, while another ran for 17 min 45 s without interruption. A serious application must keep the text already received and reconnect without losing the dictation.
4. Voxtral 3 is in public preview. Its behaviour and availability may still change: you need to be able to fall back to the previous generation without rewriting everything.

Why we chose Voxtral

• Quality on specialist French, once correction is applied, at the level we were looking for in reports.
• A European provider. Mistral AI is a French company and its API is hosted in Europe. For professionals who dictate sensitive content, that is a fundamental criterion, not a detail.
• The price, a few tenths of a cent per minute, which makes an affordable subscription possible.
• Open weights for earlier generations, which keep the door open to running it locally, on your own servers.
• A single ecosystem: correction runs on a Mistral language model, Mistral Small, the fastest of those we measured, at around 200 tokens per second.

Voxtral vs Whisper and the others

When it launched in 2025, Mistral announced that Voxtral outperformed Whisper large-v3, OpenAI’s open model; that is the vendor’s own measurement. On the independent side, Artificial Analysis ranked Voxtral Small Transcribe 3 third in its speech-to-text leaderboard when it was released. We have not rerun these comparisons: our bench measures what matters to our users, specialist terms in French.

The best model in the world is not enough, in any case, to make good professional dictation. What makes the difference day to day is everything around it: the vocabulary, the bounded correction and the memory of your corrections, explained in our article on how AI voice dictation works. And to estimate what dictation could save you, see the time-savings calculation.

Sources

• Mistral AI, “Voxtral”, July 2025.
• Mistral AI, “Voxtral Transcribe 2”, February 2026.
• Mistral AI, “Voxtral 3”, October 2026.
• Artificial Analysis, Speech to Text leaderboard.
• TechVoice internal measurements, synthetic corpus of orthopaedic reports in French, November 2026.