# Mistral AI Voxtral 3 bundles speech, TTS and voice agents in one SDK

> The Paris-based model maker enters the real-time voice stack with 50-plus-language transcription, $0.0035-per-minute pricing and EU data-control options.

*Published 2026-10-21 · By Marcus Vance*

Mistral AI Voxtral 3 is a family of speech-to-text and text-to-speech models paired with a LiveKit-based voice agent SDK that streams listening, reasoning, tool calls, and speaking as one real-time pipeline.

Mistral AI released Voxtral 3, a family of speech-to-text and text-to-speech models bundled with a voice agent SDK for real-time, multi-step conversations. The models are available in public preview through the Mistral API on EU, U.S. and global endpoints, or for self-deployment on customer infrastructure, according to the company. The release extends Mistral beyond text and reasoning models into the voice layer, an area CEO Arthur Mensch has called the most important interface to AI systems.

Voxtral 3 expands Mistral's audio line to cover three jobs: turning speech into text, turning text into speech, and running both in live conversations, according to Mistral's documentation. The new family adds streaming variants and an SDK built on LiveKit, an open-source real-time communications framework. Developers can use the same product family for batch transcription, cloned-voice synthesis and full voice-agent sessions.

## Background: Why Mistral is entering the voice stack

Voice has historically been a wrapper around text models: a speech recognizer feeds a chatbot, and a synthesizer reads the reply. Voxtral 3's SDK collapses those layers into one pipeline. Mistral is following OpenAI and ElevenLabs into realtime voice, but it is doing so with an open-source runtime and a sovereign-deployment option that keeps audio processing inside the customer's chosen region.

The release also signals a shift in how Mistral packages its technology. Instead of exposing speech models as separate endpoints and leaving developers to stitch them together, Mistral is shipping an agent developer experience. The SDK handles the streaming loop, so a developer can add voice to an application without building a custom realtime orchestration layer.

## What is Voxtral 3 and why does it matter?

Voxtral 3 is a product family rather than a single model. The lineup includes Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, Voxtral Mini Transcribe Realtime 3, Voxtral Mini TTS 3 and Voxtral Speech 3. The transcription models convert recorded files and live streams into text with speaker labels, timestamps and custom vocabulary. The TTS models generate speech with preset or cloned voices, support emotional adaptability, and accept tags for pace and emphasis.

The voice agent SDK is the differentiator. It binds listening, reasoning, tool calling and speaking into a single streaming pipeline, so an agent can hear a user, decide on an action, call a tool and respond in one continuous exchange. Mistral says the pipeline is designed for sub-second turns, which is the threshold where voice agents begin to feel conversational rather than mechanical.

- Stream audio from a microphone or telephony source into a Voxtral realtime transcription model.
- Run the transcribed text through a reasoning model or agent loop, including tool calls.
- Select a response and feed it to a Voxtral TTS model.
- Stream synthesized speech back to the user with end-of-turn detection.
- Optionally keep speaker labels, timestamps and session logs for compliance or analytics.

## How does the Voxtral 3 voice agent pipeline work?

Mistral's design places voice models close to reasoning models at inference time. Arthur Mensch, co-founder and CEO of Mistral AI, said the company deploys voice models physically close to reasoning models, a choice intended to cut the network hops that add latency to spoken interactions. The architecture matters because a voice turn is a chain of dependent operations: recognition, reasoning, tool execution and synthesis must all complete before audio reaches the user.

The realtime transcription variant targets a streaming delay of 80 to 1,200 milliseconds, according to Mistral's documentation. The company separately says realtime transcription achieves sub-500 ms latency in more than 35 languages. That range places Voxtral 3 in the same performance territory as telephony-grade speech systems, where delays above one second disrupt natural conversation.

For agent conversations, latency is not only about recognition speed. The system must detect when a user has finished speaking, decide whether a tool call is required, and phrase the reply. Voxtral 3 bundles end-of-turn detection and context-biasing with the streaming models, reducing the glue code developers previously had to write between separate speech APIs.

The SDK's LiveKit foundation gives it a transport layer designed for realtime audio and events. That matters for production voice agents, which need to handle connection drops, audio jitter and concurrent sessions. By inheriting LiveKit's infrastructure, Voxtral 3 can be embedded into telephony, browser and mobile applications without a proprietary realtime stack.

## What are the technical specifics of the Voxtral 3 model family?

The batch transcription workhorse is Voxtral Mini Transcribe 3, a 4-billion-parameter model. Mistral prices it at $0.0035 per minute of audio. The model supports more than 50 languages, including all 24 official EU languages plus low-resource Asian and Indic languages, with diarization, context-biasing and end-of-turn detection.

The text-to-speech side splits into Voxtral Mini TTS 3 and Voxtral Speech 3. Both support more than 35 languages and 50-plus dialects, with emotional adaptability and tags for pace and emphasis. Voice cloning is available from roughly 30 minutes of audio. Mistral prices TTS at $13 per 1 million characters.

*Voxtral 3 model family at a glance*

| Model | Type | Key specifications | Pricing |
| --- | --- | --- | --- |
| Voxtral Mini Transcribe 3 | Batch speech-to-text | 4B parameters; 50-plus languages | $0.0035 per minute |
| Voxtral Mini Transcribe Realtime 3 | Streaming speech-to-text | 35-plus languages; 80-1,200 ms streaming delay | API or self-deployment |
| Voxtral Mini TTS 3 | Text-to-speech | 35-plus languages; 50-plus dialects; cloned voices | $13 per 1M characters |
| Voxtral Speech 3 | Text-to-speech | 35-plus languages; emotional adaptability | $13 per 1M characters |

## How does Voxtral 3 pricing compare with rivals?

Mistral's published rates give developers a predictable cost basis for large-scale voice deployment. Batch transcription at $0.0035 per minute and TTS at $13 per 1 million characters are the two anchor prices in the release. The company did not publish a separate price for the realtime transcription variant or for the voice agent SDK.

The more significant comparison may be architectural. Developers building voice agents with separate speech APIs typically pay for recognition, synthesis and orchestration from different vendors. Voxtral 3 sells transcription, synthesis and the agent runtime as one product, which can reduce integration cost even when the per-minute price is similar to a competitor's.

## What are the market and sovereignty implications?

Voxtral 3 lands at a moment when European enterprises are pressing for AI that can run on EU infrastructure. Mistral offers API endpoints in the EU, U.S. and global regions, and it supports self-deployment on customer infrastructure. For regulated sectors, that removes a barrier: audio does not have to leave the organization or the region for processing.

Language coverage is a second sovereign-AI lever. Supporting all 24 official EU languages plus low-resource Asian and Indic languages gives public-sector customers a path to voice systems that do not depend on U.S.-hosted models for minority languages. Mistral says the supported list of languages is expanded to 50, a figure that includes languages often ignored by larger U.S. platforms.

The SDK is built on LiveKit, an open-source real-time communications framework. That choice makes Voxtral 3 portable across telephony, web and mobile deployments, and it lets customers keep orchestration inside their own infrastructure. It also signals that Mistral is targeting developers building production voice agents, not just speech-to-text pipelines.

## What did experts say?

Mistral chief executive Arthur Mensch framed the release in strategic terms. “Voice will be the most important interface to AI systems, and we deploy voice models physically close to reasoning models,” Mensch said in a post on X. The statement ties Voxtral 3 to Mistral's broader model strategy rather than positioning it as a standalone speech utility.

> Voice will be the most important interface to AI systems, and we deploy voice models physically close to reasoning models.Arthur Mensch, co-founder and CEO, Mistral AI

Mistral's release notes describe the SDK as the mechanism that makes voice agents practical. “To address this, we introduce Voxtral 3: a new family of state-of-the-art speech-to-text and text-to-speech models, alongside a voice agent SDK that brings listening, thinking, tool calling, and speaking into a single real-time streaming pipeline,” the company said.

## What's next for Mistral AI's voice strategy?

Voxtral 3 is in public preview with published pricing, a sign that Mistral expects developers to test it in production workloads. The company is moving up the stack from models to agent infrastructure, and it is doing so with voice as the front door. The combination of a LiveKit-based runtime, EU data-control options and multilingual coverage gives Mistral a differentiated bundle in a market crowded with point solutions.

The competitive field includes OpenAI's realtime voice APIs and ElevenLabs' speech platform, but Voxtral 3 pairs those capabilities with European data sovereignty and an open-source orchestration layer. If sub-500 ms latency holds under production load, Mistral could become a default choice for European voice agents and for any organization that needs broad language coverage without exporting audio.

Tool calls add a further variable. A voice agent that books a flight or queries a database will be only as fast as the tool it invokes, so the sub-second target applies to the speech loop rather than the complete task. That distinction matters for buyers comparing Voxtral 3 with text-based agent frameworks that bolt on speech afterward.

## Sources

1. [Voxtral 3 introduces a voice agent SDK and speech models for 50-plus languages, with Voxtral Mini Transcribe 3 priced at $0.0035 per minute and TTS priced at $13 per 1 million characters.](https://mistral.ai/news/voxtral-3/)
2. [Mistral audio covers speech-to-text, text-to-speech and live conversations; Voxtral 3 models transcribe recorded files and live streams with speaker labels, timestamps and custom vocabulary, with streaming delay from 80 to 1,200 ms.](https://docs.mistral.ai/capabilities/audio)
3. [Mistral CEO Arthur Mensch said voice will be the most important interface to AI systems and that the company deploys voice models physically close to reasoning models.](https://x.com/arthurmensch/status/2112589831641710797)

---
Source: https://aiintelreport.com/frontier-models/voxtral-3-mistral-ai-voice-agent-sdk
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt