Mistral Introduces Voxtral 3: 50-Language Recognition and Voice Agents for the Enterprise

France's Mistral AI has released the Voxtral 3 model family, designed to build end-to-end voice agents for enterprise environments. The release combines automatic speech recognition (STT), speech synthesis (TTS), and a specialized Voice Agent SDK. The primary focus is cutting latency below 500 ms while enabling full on-premises or EU-based cloud deployments, allowing enterprises to maintain data sovereignty without sending raw audio to US third-party providers.

Recognition Architecture and Benchmarks

Voxtral Small Transcribe 3 supports 50 languages, including all 24 official EU languages alongside less common Asian and Indian dialects. The architecture integrates speaker diarization with a context-biasing mechanism, allowing developers to inject domain-specific terminology lists to boost accuracy without fine-tuning model weights. On the multilingual FLEURS 45 benchmark, the model delivered an average word error rate (WER) of 4.0–4.1%, outperforming Whisper large-v3 (14.9%), Gemini 3.5 Transcribe (4.5%), and Scribe v2 (4.9%).

The lightweight Voxtral Mini Transcribe 3, featuring 4 billion parameters, is optimized for minimal inference cost at $0.0035 per minute of audio. For conversational agent setups, Mistral built the Voxtral Realtime Transcribe 3 variant with native end-of-turn prediction. The underlying algorithm evaluates acoustic and semantic cues simultaneously, letting virtual agents respond without artificial pauses while avoiding cutting off the speaker mid-sentence.

Context-Aware Speech Synthesis

Speech generation is handled by Voxtral Mini TTS 3.0, which supports over 35 languages and 50 dialects. The model analyzes the emotional context of the input prompt, automatically tuning prosody, rhythm, and semantic emphasis without requiring manual SSML tags. Developers can also use Pro Voice, a dedicated tool that clones a speaker's style, accent, and dynamic emotional range from 30 minutes of reference audio.

Key specifications and metrics for the Voxtral 3 line:

• Recognition accuracy: 4.0% WER on FLEURS 45 and 94.6% accuracy on key-term extraction using context prompting;
• Diarization: 18.4% cpWER across mixed AMI, Switchboard, and AISHELL-4 corpora;
• Streaming latency: sub-500 ms in Voxtral Mini Realtime 3 mode;
• Pricing: $0.0035 per minute for transcription (Mini) and $13 per 1 million characters for synthesis (TTS);
• Infrastructure: EU-hosted API endpoints or complete on-premises deployment for air-gapped environments.

By packaging an SDK with native function calling, Mistral turns Voxtral 3 into a turnkey foundation for contact centers, clinical desks, and automated customer support. In one demonstration, an agent transcribed a French dermatology referral, queried a clinic's internal API to locate an open slot, and confirmed the appointment in English. Providing direct access to model weights for local execution gives Mistral a major edge in heavily regulated European sectors where cloud-based voice APIs from OpenAI or Google face compliance roadblocks.