Voxtral 3: Mistral AI wants its voice agents to speak in under a second

Mistral AI introduces Voxtral 3, its new transcription and speech synthesis models, featuring an agent SDK, 50 languages, and sovereign deployment.

A conversation with a voice agent can still feel like talking to a sophisticated answering machine: the system waits for the end of a sentence, interprets the request, then prepares a response before speaking again. With Voxtral 3, introduced on October 20, 2026, Mistral AI aims to reduce these delays by pairing new transcription and speech synthesis models with a SDK capable of coordinating the entire conversation. The platform supports understanding complex requests, using external tools, and delivering expressive responses in dozens of languages. It can run from Mistral’s API or be deployed on a private infrastructure, an option specifically aimed at enterprises that wish to maintain control over their voice data. Understanding Fifty Languages and Recognizing Speakers

The first component of Voxtral 3 involves transcription. With Voxtral Small Transcribe 3, Mistral offers a model designed to convert speech to text, identify different speakers, and integrate vocabulary specific to a business activity. The family also includes Voxtral Mini Transcribe 3, a four-billion-parameter variant intended for use cases where cost and required processing resources are decisive factors.

Both models cover over fifty languages, including the twenty-four official languages of the European Union. Their context-biasing feature allows users to pass keywords or textual prompts to improve the recognition of proper nouns, technical terms, and uncommon expressions. A customer service department can thus provide the tool with the names of its products or employees without having to conduct new training. Diarization complements this approach by allowing the system to distinguish who spoke and at what time, a feature useful for meeting analysis and interview transcriptions.

On FLEURS, in a comparison covering forty-five languages common to seven systems, Voxtral Small Transcribe 3 shows an average error rate of 4.0% when the language is automatically detected, and 4.1% when it is specified. Mistral places it ahead of several competitors, including Whisper large-v3, Scribe v2, and Gemini 3.5 Transcribe, depending on the protocols used. The figures come from the evaluation presented by the company and should not be equated with uniform accuracy across every language, accent, or acoustic environment. Detecting the End of a Sentence Before Responding

Real-time transcription presents an additional challenge: the agent must not only understand the words, but also determine when it can respond. A pause can signal the end of a request or simply a hesitation. Reacting too early interrupts the user; waiting too long artificially slows down the conversation.

Voxtral Mini Transcribe Realtime 3 combines speech recognition with a turn-end detection system that takes into account acoustic cues and the meaning of what is being said. Mistral announces transcription in less than 500 milliseconds in certain configurations, with adjustable latency depending on the desired trade-off between speed and reliability. The service operates via streaming through WebSocket and retains the diarization and contextual vocabulary features.

This latency concerns speech processing, and not necessarily the full duration of a response generated by an assistant. Mistral's results also show that a faster configuration can increase the transcription error rate. The system must therefore strike a balance between speed, recognition quality, and the ability to let the user finish speaking. Speech Synthesis Capable of Interpreting Emotions

To produce responses, Mistral presents Voxtral Mini TTS 3, also marketed under the name Voxtral Speech 3. The model supports over thirty-five languages and fifty dialectal variants, with control over pronunciation and vocal style. Its objective is not merely to read a text correctly, but to convey the nuances that make speech believable: rhythm, emphasis, intensity, and emotional tone.

Two control modes are offered. The first relies on an implicit interpretation of the context: the model infers how to pronounce a sentence based on its meaning and the conversation. The second allows explicit prompts to be inserted into the text using tags such as `[emotion=sad]` or `[pace=slow]`. A developer can thus guide the expression of a line without modifying the overall voice being used.

Mistral also introduces Pro Voice, an adaptation feature intended for enterprise use. Using a recording of up to thirty minutes, it seeks to replicate the identity, intonations, and expressive register of a speaker across the various supported languages. This customization can be used to maintain a consistent voice identity across multiple markets, subject to having the necessary rights and consent to use the recordings.

The company highlights its performance in pronunciation fidelity, particularly regarding numbers, proper nouns, and other content likely to cause errors in speech synthesis. However, the results stem from specific protocols, some of which use another AI model as an evaluator, and do not constitute a universal measure of human preference. An SDK That Separates Listening, Reasoning, and Speech

One of the main innovations of Voxtral 3 lies in its voice agent SDK built on LiveKit. Unlike an architecture where each step waits for the completion of the previous one, the system organizes several components into a continuous pipeline. Voxtral Mini Transcribe Realtime 3 listens and determines when the user has finished, a general-purpose model like Mistral Small interprets the request and uses tools if necessary, and then Voxtral Speech 3 generates the response.

This separation retains technical flexibility. Developers can replace the reasoning model, select a different voice, or adapt the output language without rewriting the entire application. Mistral announces conversational turns executed in less than one second in its optimized configuration, with response barge-in when the user starts speaking again. This delay does not guarantee that an