Voxtral 3: Mistral's New Voice Agent SDK and State-of-the-Art Speech Models

Voxtral 3 is a new family of state-of-the-art speech-to-text and text-to-speech models, coupled with a voice agent SDK, designed to solve the complex challenge of building reliable, end-to-end voice agents that can listen, reason, interact, and handle multi-step workflows in real time.

This suite allows organizations to create natural conversational experiences—from contact center agents to hands-free copilots—while maintaining data sovereignty through deployment options on Mistral's EU-hosted API or on private infrastructure.

Core Capabilities of Voxtral 3

Voxtral 3 addresses the four critical requirements of a reliable voice agent: accurate listening, deep contextual reasoning, natural speech generation, and multi-step action execution.

Multilingual and Accurate Transcription

The Voxtral Small Transcribe 3 model provides state-of-the-art multilingual transcription, supporting over 50 languages, including 24 official EU languages and numerous low-resource Asian and Indic languages. Key features include:

  • Diarization: The model accurately identifies who spoke when, minimizing confusion in multi-speaker conversations.
  • Context Biasing: Users can provide key terms and free-form text to bias the transcription, significantly improving recall on domain-specific vocabulary without requiring model fine-tuning.
  • End-of-Turn Detection: It natively detects when a speaker has finished, allowing agents to respond with low latency and minimizing false cut-off rates.

For real-time agentic use cases, Voxtral Mini Transcribe Realtime 3 achieves high-quality transcriptions at sub-500ms latency across over 35 languages, also supporting diarization and context biasing.

Expressive and Adaptive Speech Generation

The Voxtral Mini TTS 3 model offers state-of-the-art expressiveness and instruction following across 35+ languages and 50+ dialects. The model's naturalness stems from its ability to implicitly interpret the semantic meaning of the text context to determine appropriate emotion and voice style.

Key features of the TTS model include:

  • Emotional Adaptability: Voices adjust their emotion, pace, and emphasis based on the conversational context.
  • Voice Cloning: Voxtral Pro Voice adaptation can reproduce a speaker’s distinctive character, emotional range, and speaking style across languages using only 30 minutes of audio.
  • Cost Efficiency: The model is priced at $13 per 1M characters, which the source material notes is less than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash.

The Real-Time Voice Agent Pipeline

The Voxtral 3 voice agent SDK integrates the entire conversational flow into a single, real-time streaming pipeline, ensuring minimal delay between turns.

Agentic Workflow and Latency

The agent operates by chaining separate models: Voxtral Mini Transcribe Realtime 3 listens and detects speech completion; a general-purpose LLM (such as Mistral Small) reasons and calls tools; and Voxtral Speech 3 replies in the tone of the conversation. Because no stage waits for the previous one to finish, each turn completes in under one second, and the agent stops speaking immediately if the caller interrupts.

This architecture enables agents to perform complex tasks during live conversations, such as verifying identity, checking bills, or booking appointments, by utilizing function calling, web search, and retrieval over the user's private systems.

Deployment and Sovereignty

Voxtral 3 is designed for sovereign deployment, allowing organizations to control where audio is processed and how data is handled. The models are available through Mistral's EU-hosted API or can be self-deployed on the organization's own infrastructure. This capability is a significant draw for enterprises requiring strict data control.

Performance and Market Insights

Voxtral 3 demonstrates competitive performance across several industry benchmarks, particularly in multilingual and real-time scenarios.

Benchmark Performance

  • Transcription: Voxtral Small Transcribe 3 outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe in multilingual transcription and diarization benchmarks.
  • Real-Time Transcription: Voxtral Mini Transcribe Realtime 3 is evaluated against the FLEURS 35 benchmark for real-time Word Error Rate (WER).
  • TTS Quality: Internal evaluations show Voxtral TTS maintains a high quality-per-dollar ratio across metrics like Voice Code Bench (handling numbers and unique names) and Pronunciation Control.

Community Discussion and Use Cases

Community feedback highlights both the advanced capabilities and the practical limitations of current AI voice technology:

  • Enterprise Integration: The SDK is viewed as highly valuable for automating complex customer service workflows. One user noted the potential for agents to interact directly with company intranets to handle requests like payment authorization.
  • Sovereignty vs. Openness: A recurring theme in the discussion is the tension between the need for data sovereignty (which Voxtral 3 addresses) and the preference for open-source models, with some users expressing disappointment that the models are closed.
  • Voice Quality Perception: Some users noted a perceived shift in voice quality, with some finding the newer, more expressive voices to sound more artificial compared to older versions.

Pricing and Availability

Voxtral 3 is available today via the Mistral API. Pricing varies by model and usage type:

Model Usage Type Price
Voxtral Mini Transcribe 3 Batch $0.0035/min
Voxtral Mini Transcribe Realtime 3 Realtime $0.008/min
Voxtral Small Transcribe 3 Batch $0.0040/min
Voxtral Speech 3 Batch and Realtime $13/1M chars

Sources

Related

  • Dispatch
    Mistral AI Voxtral 3 release notes

    Mistral AI has launched Voxtral 3, a suite of state-of-the-art speech-to-text and text-to-speech models alongside a voice agent SDK designed for low-latency, multi-step conversational automation.

  • Dispatch
    Mistral AI Voxtral TTS Release

    Mistral AI has released Voxtral TTS, a 4B-parameter multilingual text-to-speech model that provides high-naturalness voice generation and low-latency streaming for enterprise voice agents.

  • Dispatch
    Mistral AI Voxtral Release

    Mistral AI has released Voxtral, a family of open-weight speech understanding models in 24B and 3B sizes that provide state-of-the-art transcription and native semantic understanding under the Apache 2.0 license.

  • Dispatch
    Mistral AI Voxtral Transcribe 2 Release

    Mistral AI has released Voxtral Transcribe 2, featuring Voxtral Mini Transcribe V2 for high-efficiency batch processing and Voxtral Realtime for ultra-low latency live transcription.

  • Dispatch
    Mistral AI Devstral Release

    Mistral AI has released Devstral, an agentic LLM for software engineering that outperforms open-source models on the SWE-Bench Verified benchmark and is available under the Apache 2.0 license.