Voxtral 3 voice agent SDK: 50+ languages and emotional speech

Mistral has announced Voxtral 3, a voice agent SDK and speech model family designed to handle real-time, multi-turn conversations across 50+ languages. The toolkit addresses a core challenge in voice AI: building agents that can listen accurately, reason about intent, speak naturally, and execute tool calls simultaneously.

What changed

Mistral introduced Voxtral 3 as a comprehensive voice agent system. The announcement includes three main components: Voxtral Small Transcribe 3 (a speech-to-text model), Voxtral Mini TTS 3 (a text-to-speech model with emotional adaptation), and a new voice agent SDK that integrates listening, thinking, tool calling, and speaking into a single real-time pipeline.

The transcription model supports 50+ languages, including 24 official EU languages and low-resource Asian and Indic languages. The company reports that the Small variant achieves state-of-the-art results in diarization (identifying who spoke when) and multilingual transcription, outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe. A 4-billion-parameter Mini variant is available at $0.0035 per minute for faster, lower-cost deployments.

Voxtral Realtime Transcribe 3 is a transcription model built for agent use cases, achieving high-quality transcriptions at sub-500ms latency across over 35 languages, including diarization and context biasing. The model natively supports end-of-turn detection, using acoustic and semantic information to predict when a user has finished speaking.

The text-to-speech component, Voxtral Mini TTS 3, operates across 35+ languages and 50+ dialects. At $13 per million characters, the company positions it as less expensive than OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. The model generates speech with contextual awareness of emotion, adjusting pace and emphasis based on conversational content.

For voice adaptation, Mistral introduced Pro Voice, which reproduces a speaker’s distinctive character and emotional range using 30 minutes of audio, working across multiple languages.

The voice agent SDK orchestrates these models to handle multi-step workflows in real time. A reference scenario demonstrates a French-language medical referral: the agent transcribes the user’s speech, retrieves a referral letter, calls a tool to book an appointment, and responds with contextual inflection.

Deployment flexibility is a core feature. Organizations can use Mistral’s EU-hosted API or self-deploy Voxtral 3 on their own infrastructure, maintaining control over where audio is processed and whether conversations are retained.

Why it matters

Voice agents remain difficult to build end-to-end because they require simultaneous accuracy in four critical areas. Mistral’s bundled approach, combining transcription, reasoning-aware text-to-speech, and deep tool integration, offers an alternative to assembling components from multiple vendors. This is particularly relevant for professionals building contact-center systems, service desks, hands-free copilots, or multilingual virtual assistants.

The language breadth matters significantly. Support for 50+ languages, including low-resource variants, expands voice agents beyond English-dominant use cases. The inclusion of 24 EU languages signals focus on regulatory and operational compliance in a region with strict data residency requirements.

Pricing clarity helps cost planning. At $0.0035 per minute for transcription and $13 per million characters for speech synthesis, the models provide explicit benchmarks. Teams evaluating voice infrastructure can compare these costs against other agentic deployments.

Pro Voice adaptation addresses a genuine pain point. Voice reproduction from 30 minutes of audio suggests higher-fidelity, professional-grade output compared to basic voice cloning methods. This matters in scenarios where brand voice consistency or speaker authenticity is critical.

The sovereignty model, support for self-deployment and EU hosting, appeals to organizations bound by GDPR or internal data policies. Keeping audio processing within your infrastructure or Mistral’s EU region removes intermediaries and provides control over data handling.

What to test

Before adopting Voxtral 3, teams should validate several vendor claims and practical constraints:

• Transcription accuracy under realistic conditions. Test diarization, end-of-turn detection, and context biasing on domain-specific vocabulary within your actual use case, whether customer support calls, medical consultations, or other scenarios. Performance varies by language pair, accent, and audio quality.

• Real-time latency and throughput. Verify that the sub-500ms latency claimed for Voxtral Realtime Transcribe 3 holds in your deployment context. Test with concurrent conversations to understand how latency scales under load.

• Tool calling and reasoning integration. The SDK promises deep tool use, but confirm that orchestration handles interruptions, context switches, and multi-step workflows without hallucination or dropped steps. Simulate edge cases where the user changes intent mid-conversation.

• Emotional speech quality and consistency. Listen to speech samples across languages and dialects. Verify that emotion, pacing, and emphasis variables produce natural, contextually appropriate output rather than mechanical variation.

• Pro Voice quality on your audio. If using Pro Voice adaptation, test whether the 30-minute audio requirement produces accurate reproduction of the speaker’s emotional range, accent, and distinctive traits across languages.

• Deployment mechanics and cost scaling. For self-deployed instances, benchmark infrastructure cost, latency, and throughput. For API-based deployments, confirm pricing applies as described under your expected call volume and language mix.

• Fallback and error handling. Test system behavior when transcription fails, tool calls time out, or the user speaks a language outside the supported set.

Common questions

What languages does Voxtral 3 support?

Voxtral 3 supports transcription and speech across 50+ languages, including 24 official EU languages and low-resource Asian and Indic languages. The text-to-speech component operates in 35+ languages and 50+ dialects.

How much does Voxtral 3 cost?

The Mini transcription model costs $0.0035 per minute. The Mini TTS 3 model costs $13 per 1 million characters. These rates are positioned as lower than OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash.

Can I deploy Voxtral 3 on my own infrastructure?

Yes. Mistral offers Voxtral 3 via an EU-hosted API or for self-deployment on your own infrastructure. This allows organizations to retain control over audio processing, data handling, and conversation retention.

How does Voxtral 3 differ from OpenAI’s voice capabilities?

The announcement does not directly compare Voxtral 3 to OpenAI’s voice offerings. However, Voxtral 3 emphasizes broader language support (50+ languages), emotional speech adaptation, deep tool integration, and sovereign deployment options. Teams should run comparative tests on specific metrics like transcription accuracy, latency, and cost for their use case.

The conclusion

Voxtral 3 represents an end-to-end voice agent platform built for professionals who need multilingual, emotionally aware, real-time conversations with tool access. The breadth of language support, explicit pricing, and sovereignty options differentiate it from point solutions and closed ecosystems.

What remains to validate is whether claimed state-of-the-art performance translates to your specific domain, accent profile, and tool ecosystem. The reference example, a French medical referral scenario, is polished; production deployments across contact centers, low-resource languages, or challenging audio conditions will provide the real measure of performance.

Watch for how quickly Pro Voice adoption scales. If the 30-minute audio requirement becomes standard for professional voice work, it signals a shift toward higher-fidelity voice adaptation. Also monitor whether other agentic platforms adopt Voxtral 3 models as a base offering, which would validate its positioning as an alternative to OpenAI-centric voice stacks.

Related reading

• MiniMax M3.1 Flash: $0.15/M tokens cuts agent costs 70%
• Harness2 framework improves agent performance 11.5 points
• Nemotron 3.5 Super VL 120B: Open Multimodal Agent Model