Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵

Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription, diarization and context-biasing. It detects when someone has finished speaking so agents can respond without delay.

Voxtral Mini TTS 3 achieves state-of-the-art expressiveness in 35+ languages and 50+ dialects. Implicit instruction-following reads emotion, pace, and emphasis from the conversation. With explicit instruction-following, tags like [emotion=sad][pace=slow] set them line by line.

With Pro Voice, we’re also rolling out professional voice adaptation. Record up to 30 minutes of audio to replicate your voice with extreme fidelity across every supported language.

Voxtral 3 also introduces a voice agent SDK, built on LiveKit. Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation.

More here:

mistral just keeps cooking quietly

Europe is back 🔥 Voxtral 3 is SOTA multilingual transcription over 50+ languages, TTS at $13/1M chars, and a voice agent SDK on LiveKit One pipeline for listen, reason and speak is where voice agents are headed. My guess: EU-hosted or self-deployed is the enterprise pitch Excited to compare this with others 👀

35+ languages and 50+ dialects is a lot

Mistral should launch an Android keyboard that has voice input. I would pay to get rid of Google

are the weights open on this one, for anything with customer calls in it thats the first question a business will ask

One voice pipeline, zero margin for error. Consent and latency decide if it ships.

the real benchmark is if voxtral 3 can transcribe a crypto space with five people talking over each other

Mistral just priced transcribing an hour of audio at 24 cents. ElevenLabs has a pricing problem now.

But why no open weight models?

@qtnx_ Wow it’s fast. ChatGPT killer from the EU!

the two of them sitting there like a real dinner party while its just a robot on the other end is honestly the funniest part of this whole rollout

it's crazy that nobody on this planet has come even close to what Sesame did over a year ago

Why AssemblyAI isn't part of the benchmarks here?

Voxtral 3's voice agents API looks promising - the explicit instruction tags like [emotion] and [pace] are a clever touch. What kind of latency are you seeing on the transcription pipeline in practice?

No hugging face link ?

Line-by-line emotion tags are the real upgrade. [emotion=sad][pace=slow] means I can direct a game character's voice like a script instead of hoping the model guesses. Going to try it on an NPC in one of my web games, in Spanish.

Measured latency ?

✅ Worth a look if you build voice apps. $0.0035/min transcription and $13 per 1M characters of speech is cheap, and swapping the voice, language or model without rebuilding the agent is the practical win. Test it on real, noisy phone audio before you switch.

Voxtral 3 sounds like the next step for voice agents. The real test will be how well this handles heavy background noise in messy environments.

Please add my voice as a Sarcastic frenchman

every transcript of me is just uhh

voice agents are about to make 'please hold' the most expensive phrase in customer service

One API for a fully autonomous real-time voice agent!

Let me think ….. Nope

The turn-detection detail matters here: Voxtral Small Transcribe 3 detects when someone has finished speaking so agents can respond without delay. Combined with Mini TTS 3 in 35+ languages and 50+ dialects, that covers both sides of a voice conversation.

Excellent

Does it support Greek?

Transcription accuracy stopped being the bottleneck for voice agents a while ago. The hard part is turn-taking: knowing when someone has actually finished, and stopping mid-sentence when they cut in. Curious what end-to-end latency the API hits with interruptions on.

Voxtral 3 for voice agents sounds great, how's the latency?