Mistral just shipped Voxtral 3, and it might be the most useful voice AI release of the month. While everyone argues about chatbots, this French lab rolled out speech models that listen, transcribe, and talk back in more than 50 languages. Here’s what it actually does, what it costs, and whether you should care.
What Voxtral 3 actually is
Voxtral 3 isn’t one model. It’s a whole family of speech tools, and each piece handles a different job.
First, there’s the transcription side. Voxtral Small Transcribe 3 turns audio into text in over 50 languages, including all 24 official EU languages plus a bunch of low-resource Asian and Indic languages that most services ignore. It also does diarization, which is a fancy word for figuring out who said what in a conversation. That matters if you ever tried to read a transcript of a three-person meeting where every line got blamed on the wrong person.
Then there’s the speaking side. Voxtral Mini TTS 3 reads text out loud in 35+ languages and more than 50 dialects, and it adjusts emotion, pace, and emphasis based on what the text actually means. No special voice instructions needed. The model reads “ugh, great” differently from “great, thanks!” on its own.
Finally, Mistral wrapped everything in a voice agent SDK built on LiveKit. That’s the piece that turns all of this into something you can actually build with, which we’ll get to in a minute.
The four models and what they cost
Pricing on Mistral’s API is pay-per-use, and honestly, the numbers are low enough to raise an eyebrow. Here’s the full breakdown straight from the launch page:
| Model | What it does | Price |
|---|---|---|
| Voxtral Small Transcribe 3 | Top-tier transcription, 50+ languages | $0.004/min |
| Voxtral Mini Transcribe 3 | Smaller, 4B params, batch jobs | $0.0035/min |
| Voxtral Mini Transcribe Realtime 3 | Live transcription under 500ms | $0.008/min |
| Voxtral Speech 3 | Text to speech, 35+ languages | $13 per 1M chars |
To put that in perspective: transcribing an hour of audio on the Mini model costs about 21 cents. A single hour of human transcription service runs $60 or more. Meanwhile, generating an hour of spoken audio costs roughly 47 cents at typical reading speeds.
If you want to run it yourself instead, that’s allowed too. Everything ships for self-deployment on your own servers, and Mistral hosts the API in Europe if you’d rather not manage hardware at all.
What it beats
Benchmarks are easy to cherry-pick, so treat this with the usual salt. Still, the numbers here are worth a look.
On the FLEURS multilingual benchmark, Voxtral Small Transcribe 3 outperformed OpenAI’s Whisper, ElevenLabs’ Scribe v2, and Gemini 3.5 Transcribe across the 45 languages all seven tested models support. It also led on diarization tests and on keyphrase recall, which measures whether the model catches specialized vocabulary you feed it.
That last feature is called context biasing, and it’s quietly one of the best reasons to pick this over free alternatives. You hand the model a list of product names, people’s surnames, or jargon before transcribing. It then uses that list to avoid mangling exactly the words that matter most. Podcast about niche medical devices? Feed it the device names. Client calls full of acronyms? Same trick.
So is it a Whisper killer? For multilingual work with domain vocabulary, possibly. For a quick English-only voice memo, Whisper running free on your laptop still does the job fine.
What beginners can actually build
This is where it gets fun. Because the models are cheap and the SDK wires them together, a few practical projects come to mind for regular people, not just engineering teams.
- A podcast transcriber that knows the names. Drop in audio, supply the guest names and niche terms, and get clean transcripts with speaker labels for pennies.
- A voicemail or call-in summarizer for a small business. Missed calls become searchable text, with who-said-what intact.
- Audiobook-style voiceovers for your own writing. Blog posts, newsletters, or course scripts become natural-sounding audio in dozens of languages, for cents.
- A simple phone agent. The SDK handles listening, thinking, and speaking as one streaming pipeline, so a basic “answer questions about my store hours” bot is a weekend project rather than a research lab.
If you already pay for tools that do pieces of this, you have options. We covered how Gemini’s text to speech works for script reading, and there’s a guide on turning a PDF into an audiobook with AI you already pay for. Voxtral 3 competes mostly on price per minute and language coverage, so check those against what you use now.
The voice agent SDK is the real story
Transcription is a solved-ish problem. Agents that talk are not, and that’s the bet Mistral is making here.
The SDK chains four steps into one real-time stream: a transcription model listens and detects when you’ve finished talking, a general model like Mistral Small reasons about what you meant and picks tools, the system executes those tools, and Voxtral Speech 3 answers out loud in the tone of the conversation so far. Each stage is a separate model, so you can swap the brain or the voice without rebuilding everything.
There’s a twist worth knowing about. Because the models detect end-of-turn (when a speaker is actually done), responses come without that awkward lag older voice bots had. Anyone who has been interrupted mid-sentence by a robotic “How can I help you today?” knows exactly why this matters.
Mistral also pushed a professional voice option called Pro Voice. Give it 30 minutes of sample audio and it reproduces a speaker’s distinctive character and emotional range across languages. That’s either a fantastic tool for creators or a slightly unsettling deepfake kit, depending on your level of optimism. Both readings are probably correct.
How to try it today
Everything lives on Mistral’s launch page and their developer docs. For the non-technical path, here’s the short version.
- Create an account on Mistral’s platform and grab an API key.
- Pick your model: Mini Transcribe 3 for cheap batch transcription, the Realtime model for live audio, Speech 3 for voice output.
- Test in the playground first. Upload a noisy audio file with some tricky vocabulary, then try again with context biasing turned on and watch the difference.
- Only then write code. The docs include quickstart examples for transcription and speech that run in a few lines.
One caution: pricing is per minute or per million characters, so a runaway loop hammering the API does cost real money. Set a spending cap before you experiment. That advice ages well with every cloud service, by the way.
Where it fits against the competition
We compared Grok’s Voice Transcribe 2.0 against Whisper’s API pricing before, and the pattern keeps repeating: speech AI keeps getting cheaper while quality creeps up. Voxtral 3 continues that trend and adds the agent layer most rivals don’t bundle.
The catch is ecosystem. Whisper has years of tooling around it. ElevenLabs owns the “nice voices” mindshare. Google bundles speech into products you might already pay for. Mistral’s edge is raw price, open deployment, and the EU-hosted option that some European businesses legally need. If none of those matter to you, existing tools remain fine.
The takeaway
Voxtral 3 makes voice AI cheap enough to waste, which is exactly when a technology gets interesting. Try the transcription with context biasing on a messy real-world audio file, and compare the price against whatever you use now. If the numbers hold up outside the benchmark, your next voice project just got a lot less expensive.
