Mistral AI launches Voxtral 3 for real-time voice agents

Mistral AI has launched Voxtral 3, a new family of language models aimed at real-time voice agents, according to The Decoder. The release matters for teams building phone, support, booking or back-office voice apps because Mistral is packaging listening, reasoning, tool use and speech output into one streaming stack.

For creators, marketers, freelancers and small businesses, the bigger signal is that voice agents are moving beyond simple dictation and generated narration. Mistral is positioning Voxtral 3 as infrastructure for interactive conversations where the agent can understand a caller, take action in another system and answer out loud without breaking the live flow.

What Mistral is bundling in Voxtral 3

Voxtral 3 is not described as a single model. Mistral says the system is made up of a speech-to-text model, a text-to-speech model and a voice agent SDK built on LiveKit.

The named components are:

• Voxtral Small Transcribe 3 for speech recognition
• Voxtral Mini TTS 3 for speech output
• A LiveKit-based SDK for building the voice agent layer

According to the company, these pieces work together in a streaming pipeline that handles listening, thinking, external tool use and speaking. That end-to-end framing is important because many voice products require teams to stitch together separate transcription, language model, audio playback and workflow services.

Mistral says transcription supports more than 50 languages. Its text-to-speech output supports more than 35 languages and 50 dialects, with emotional adaptation. The company also says agents can connect to outside systems during live calls, such as for appointment booking or invoice checks.

That makes Voxtral 3 relevant to builders of customer-facing agents, not just teams producing voiceovers. If the agent can use tools mid-conversation, the target use case is closer to a live assistant than a passive audio generator.

Why control and residency are part of the pitch

Mistral says Voxtral 3 can be run on a customer’s own infrastructure or accessed through EU-hosted APIs. The company frames this as a way to keep data under user control.

That could matter for organizations handling sensitive conversations: client calls, support tickets, payment-related questions, healthcare-adjacent scheduling or internal operations. The source does not provide compliance claims or certifications, so buyers should not infer them. But self-hosting and EU-hosted API options give technical teams more deployment choices than a cloud-only voice service.

For small businesses, the practical question is whether they have the engineering capacity to use such a stack directly. Voxtral 3 appears most immediately relevant to developers, agencies and SaaS teams building voice features into their own products. Non-technical teams may still prefer packaged voice apps unless an integrator builds on top of Mistral’s stack.

This control angle also fits a broader pattern in agent infrastructure, where businesses want more oversight over how automated systems access tools and act on user requests. As Emonarc has covered, agent activity can run into real-world barriers when it interacts with websites and services, from shopping flows to booking tasks: Websites push back on AI agents as booking and shopping hit blocks.

How this differs from basic voice generation

Many teams first encounter AI voice through narration, dubbing or synthetic presenter workflows. Voxtral 3 is aimed at a different layer: real-time conversation.

That distinction matters. A voice generator can turn text into audio, but a voice agent needs to continuously process speech, decide what to do, possibly call an external system and respond in a way that keeps the exchange moving. Mistral’s launch combines the transcription and speech output pieces with an SDK for that agent loop.

For readers evaluating the broader voice software market, our guide to the best AI voice generators is a useful starting point for narration and production tools. Voxtral 3 sits closer to the infrastructure side, where developers build applications such as call agents, conversational ordering systems or internal voice assistants.

The category is also expanding into physical and point-of-sale settings. For example, Square recently introduced an AI ordering kiosk that customers can talk to: Square launches an AI ordering kiosk that customers can talk to. Mistral’s announcement points to the model-and-SDK layer that could power similar conversational experiences, though the source does not connect Voxtral 3 to Square.

What buyers should verify before building

Mistral has disclosed a starting transcription price of $0.0035 per minute. The source does not list broader pricing for text-to-speech, agent usage, hosting, support or enterprise terms, so teams should check the documentation before estimating total costs.

Several other due-diligence questions remain open from the public details in the source. Teams should ask about latency in their target environment, audio quality across languages, dialect coverage for their customer base, tool-calling safeguards, deployment requirements and logging controls.

The launch gives developers a new option for building live voice agents with more control over hosting. But the best fit will depend on whether a team wants raw infrastructure, has developers who can wire it into business systems and needs the data-control options Mistral is highlighting.