Mistral releases Voxtral 3 speech models for transcription and voice agents

Voxtral 3 is Mistral's new family of speech-to-text and text-to-speech models, aimed at live voice agents. A real-time model streams transcripts with latency you can set from 80 milliseconds.

Published October 25, 2026 8:09am EDT | Updated October 25, 2026 10:27am EDT · 2 min read

Add ScoviAI on Google

Mistral AI cover image (Mistral AI)· 2 min read

Key points

Mistral released Voxtral 3 on October 20, 2026, in public preview
It covers speech-to-text, real-time transcription and text-to-speech
The real-time model's latency can be set from 80 ms to 2,400 ms
Voxtral TTS 3 speaks 9 languages, with about 90 ms to the first audio

French AI company Mistral has released Voxtral 3, the next generation of its speech models. The family turns speech into text and text into speech, and Mistral is aiming it squarely at voice agents: AI assistants you talk to on the phone, in an app or in a car.

All the models are in public preview, so features and prices can still change.

What's in Voxtral 3

Voxtral Mini Transcribe Realtime 3 streams a live transcript while someone is still speaking. Mistral says you can set the delay from 80 milliseconds to 2,400 milliseconds, trading speed for accuracy. It also detects when a speaker has finished a turn, and it can tell different speakers apart.
Voxtral Small Transcribe 3 is Mistral's main model for transcribing recordings and labelling who said what.
Voxtral TTS 3 turns text into speech in 9 languages, starts speaking in about 90 milliseconds, and supports tags for 34 emotions and different speaking paces.

Mistral's own pages count the models differently (four in the changelog, five in the audio docs), so expect the lineup to settle as the preview goes on.

Why voice agents need this

A voice assistant feels natural only if it answers fast and knows when you've stopped talking. Too slow, and people talk over it; too eager, and it cuts them off. Built-in end-of-turn detection and adjustable latency are aimed exactly at that problem.

How it compares

Voice AI is crowded. ElevenLabs is the best-known name for lifelike voices, and OpenAI and Google both offer speech in their APIs. Mistral's pitch is a full set of speech tools from one European provider. If you're comparing voice tools, our guide to ElevenLabs alternatives covers the main options and prices.

What it means for you

Developers building phone or in-app assistants can test Voxtral 3 in preview now.
Everyone else may hear these voices in customer service lines and apps over the next months.

Sources

Mistral changelog (Mistral AI)
Audio overview (Mistral docs)
Voxtral TTS 3 (Mistral docs)

Source: Mistral AI

ScoviAI Team is the ScoviAI newsroom. We write every article ourselves from the original announcements and link our sources. Spotted a mistake? Tell us and we'll fix it.