In an official product announcement published on mistral.ai on October 20, 2026, French artificial intelligence laboratory Mistral AI introduced Voxtral 3, an end-to-end voice technology suite combining high-accuracy speech transcription, expressive voice synthesis, and a streaming voice agent development kit. Engineered specifically for enterprise contact centers, automated service desks, and real-time interactive copilots, the platform supports over 50 languages and is offered under sovereign deployment models either via Mistral's EU cloud or on private corporate hardware.
A central component of the release is Voxtral Small Transcribe 3 and its lightweight companion, the 4-billion-parameter Voxtral Mini Transcribe 3. In third-party evaluations across the 45-language FLEURS benchmark compiled by Artificial Analysis, Voxtral Small recorded macro-average word error rates (WER) of 4.1% in forced-language mode and 4.0% in inferred-language mode, outperforming Whisper large-v3 (14.9%) and Gemini 3.5 Transcribe (4.5%). Both models incorporate context biasing, enabling enterprise developers to feed domain-specific vocabularies, product catalogs, and proper nouns into the decoding pipeline to boost recall without requiring model fine-tuning.
In speaker diarization—identifying who spoke when in multi-party conversations—Voxtral demonstrated substantial architectural gains. Evaluated across AMI, Switchboard, and AISHELL-4 benchmarks using Concatenated Minimum Permutation WER (cpWER), Voxtral Small scored 18.4% and Voxtral Mini scored 18.6%, besting Pyannote Orchestrator Precision 3 (19.0%), Microsoft AI Transcribe 2 (20.3%), and Gemini 3.5 Transcribe (38.4%). The models also integrate native end-of-turn detection, sensing conversational pauses to trigger agent turns without artificial latency buffers.
Complementing transcription, Mistral rolled out Voxtral Mini TTS 3 for speech generation. The text-to-speech engine supports 35 languages and over 50 regional dialects, automatically adjusting prosody, speaking rate, and emotional emphasis based on the conversational tone of the underlying prompt. For bespoke branding applications, Mistral unveiled Pro Voice adaptation, an audio fine-tuning capability that captures an individual's distinctive vocal timbre, cadence, and expressive range across multiple languages using approximately 30 minutes of clean training audio.
The bundled Voice Agent SDK stitches audio input, LLM reasoning, programmatic tool calls, and audio synthesis into a unified bidirectional WebSocket pipeline. During live phone or web sessions, agents can query corporate databases, invoke REST endpoints, and stream spoken answers with sub-second turnaround. The Mini Transcribe service is priced at $0.0035 per audio minute on Mistral's EU platform, with containerized model weights available immediately for enterprise on-premise licensing.