Mistral AI has introduced Voxtral 3, a family of models and tools designed to let voice agents hear a caller, respond, and take actions while a conversation is underway. Its components are built to work together as a streaming pipeline, linking transcription, language-model processing, tool use and spoken replies.
The announcement positions Voxtral 3 as a complete voice agent stack rather than a standalone speech model. That design could let teams build conversational systems that do more than answer questions: agents can also reach external services during a live exchange.
Three components support the conversation
Voxtral 3 includes Voxtral Small Transcribe 3 for speech-to-text, Voxtral Mini TTS 3 for text-to-speech, and a voice agent SDK based on LiveKit. In combination, those parts provide the path from incoming speech to an agent’s spoken response.
Transcription covers more than 50 languages. For generated speech, the text-to-speech model supports over 35 languages and 50 dialects, with emotional adaptation. These capabilities give developers options for both understanding what a speaker says and choosing how an agent sounds in reply.
The streaming setup is central to the product’s purpose. Listening, processing, using tools and speaking are connected in one pipeline, so the agent can act within the flow of a conversation instead of treating speech recognition and response generation as separate tasks.
Agents can reach outside systems
Voxtral 3 is intended to support live interactions with external systems. Mistral gives booking appointments and checking invoices as examples of what agents can do during a conversation.
That connection changes the role of a voice agent. A system that can only transcribe and speak can relay information, while access to external tools can let it carry out steps relevant to the caller’s request. The source describes the capability at a high level; the specific services and integrations will depend on how a team builds with the SDK.
For businesses, combining speech and tool access in a single voice workflow may make it possible to handle routine requests through conversation. The key practical question is how developers connect an agent to the systems it needs and define which actions it can take.
Deployment and pricing choices
Mistral says Voxtral 3 can run on a customer’s own infrastructure or through EU-hosted APIs. The company presents these options as ways to keep data under the customer’s control.
The pricing information in the announcement starts at $0.0035 per minute for transcription. That figure applies to transcription; the source does not give pricing for the other components or the overall cost of running a complete voice agent.
Teams evaluating the stack will need to consider how the deployment option fits their requirements, alongside the transcription, speech generation and tool use their applications need. The available language coverage may also matter for products serving multilingual callers.
A connected foundation for voice products
Voxtral 3 brings several building blocks for voice applications together: speech recognition, generated speech, a LiveKit-based SDK and the ability to use external systems during calls. Rather than focusing on a single model function, the offer is aimed at developers creating agents that can participate in ongoing conversations.
The announcement leaves implementation details to the documentation. Still, its outline makes the intended workflow clear: a caller speaks, the system transcribes and processes the request, an agent can consult a connected service, and the answer can be spoken back. How useful that feels in practice will depend on the services developers connect and the conversations they design the agents to handle.