Skip to content

Mistral ships Voxtral 3 speech models and a LiveKit voice-agent SDK

Four new speech-to-text and text-to-speech models, priced from $0.0035 a minute, plus a Voice Agent SDK that chains listening, reasoning and speaking into one pipeline.

Source: Mistral AI

Mistral released Voxtral 3 on 20 October: three speech-to-text models, Voxtral Mini Transcribe 3, Voxtral Small Transcribe 3 and Voxtral Mini Transcribe Realtime 3, plus a text-to-speech model, Voxtral Speech 3, in both batch and realtime modes. Mistral says Small Transcribe 3 gets a 4.1% word error rate across 45 languages, against 16.0% for OpenAI’s Whisper large-v3, and that it beats ElevenLabs’ Scribe v2 and Google’s Gemini 3.5 Transcribe on the same multilingual tasks. The realtime model covers 40-plus languages at sub-500ms latency.

Pricing starts at $0.0035 a minute for Mini Transcribe 3, rising to $0.008 a minute for the realtime model, with speech output at $13 per million characters. Alongside the models, Mistral shipped a Voice Agent SDK built on LiveKit that chains listening, reasoning, tool calling and speaking into one streaming pipeline, with speaker diarisation, voice cloning from 30-plus minutes of reference audio, and SIP and WebRTC support for taking real phone calls.

Why it matters: building a voice agent has mostly meant stitching together a separate speech-to-text model, an LLM and a text-to-speech model yourself. An SDK that handles that pipeline, with telephony built in, removes a chunk of the glue code that usually goes into a voice product.

The caveat: the word error rate comparison against Whisper, Scribe v2 and Gemini 3.5 Transcribe is Mistral’s own benchmark, not an independent one.

Share

Related reading

ML for Web Devs

Machine learning for web developers: the only mental model you need

You do not need calculus to work with models. You need one idea, that a model is a function you fit instead of write, and the rest follows from things you already know.

ML for Web Devs

What an embedding actually is

An embedding is a list of numbers that puts similar things near each other. That one sentence is enough to build search, recommendations and retrieval. Here is how.

ML for Web Devs

Tokens, context windows, and why your bill is what it is

The unit of billing for language models is not the request or the word. Once you understand tokens and how a context window fills up, the invoice stops being a surprise.

More from The Wire

All briefs