Voxtral 3 brings Mistral AI into real-time voice

Mistral AI launches Voxtral 3, a complete stack for voice agents capable of listening, transcribing, reasoning, and responding live. The suite targets enterprise use cases, with hosting on private infrastructure or via APIs hosted in the EU.

Mistral AI has launched Voxtral 3, a suite of models and tools designed to run voice agents in real time. The package combines transcription, speech synthesis, and a voice agent SDK based on LiveKit, supporting over 50 languages for recognition and more than 35 languages plus 50 dialects for output voice. The goal is clear: to automate useful interactions, such as appointment scheduling or invoice verification, without breaking the flow of conversation.

Une pile vocale complète, pas un simple modèle de plus

Voxtral 3 is not limited to a speech recognition engine. Mistral AI pairs it with Voxtral Small Transcribe 3 for speech-to-text, Voxtral Mini TTS 3 for text-to-speech, and a voice agent SDK to orchestrate the whole. This approach chains listening, reasoning, tool usage, and oral output in a single flow, reducing back-and-forth between software components. In short, the agent doesn’t wait for the end of the sentence to start thinking. It’s used to speaking at the same time as you, which is convenient—until it interrupts you.

The positioning is as technical as it is strategic. By combining several components instead of a single generalist model, Mistral AI targets professional use cases where latency, transcription quality, and stability matter as much as the ability to respond.

Des usages d’entreprise, avec un contrôle des données revendiqué

Mistral states that the system can operate on the client’s infrastructure or via APIs hosted within the European Union. This promise addresses a sensitive point for companies: retaining control over voice data, often more sensitive than regular text. In sectors where phone exchanges contain personal information, this option weighs heavily in purchasing decisions.

The documentation cited by the company also mentions a starting price of 0.0035 dollars per minute for transcription. At this stage, comparison with other offerings will depend on volume, languages used, and hosting mode chosen [to be verified]. Still, Mistral bets on an offer that can fit into concrete business processes, from customer support to administrative management, without forcing all streams to be sent to a single platform.

Pourquoi ce lancement compte pour le marché

The move reflects a broader trend: major AI players no longer sell just models but complete ready-to-use chains. In voice, this changes the game, as value no longer depends solely on recognition quality. It also depends on latency, integration with third-party tools, voice quality, and hosting framework.

For Mistral AI, the interest is twofold. On one hand, the group strengthens its offer against American platforms already present in voice agents and automation. On the other, it pushes a European argument based on data control and local or regional execution. This is no small matter in a market where companies want assistants that speak quickly but, above all, assistants that know how to behave.

Ce que l’on sait des capacités annoncées

According to Mistral, Voxtral 3 supports over 50 languages in transcription, more than 35 languages in voice output, and 50 dialects, with emotional adaptation. The system can also call external tools during a conversation, allowing it to perform actions live rather than just respond. This type of function transforms a voice model into an agent, in the operational sense of the term.

However, these remain capabilities announced by the publisher, and their real scope will depend on measured performance in production, costs, and integration with client environments. On this front, Mistral’s suite will have to prove that real-time voice can move from demonstrator to durable work tool.

By the numbers

  • Over 50 languages — transcription, according to Mistral AI, 2026.
  • More than 35 languages — speech synthesis, according to Mistral AI, 2026.
  • 50 dialects — output voice with emotional adaptation, according to Mistral AI.
  • 0.0035 dollars per minute — starting price for transcription, according to cited documentation.

To read