Voxtral 3

Mistral AI has released Voxtral 3, a new family of speech-to-text and text-to-speech models paired with a voice agent SDK. The company announced the launch with immediate availability through its API, hosted on European, US and global endpoints, or for self-deployment on customer infrastructure. The release targets contact centers, service desks, virtual assistants and hands-free copilots.
The family includes three transcription models. Voxtral Small Transcribe 3 is pitched as the highest-accuracy option, with multilingual transcription and diarization across 50+ languages, including 24 official EU languages and low-resource Asian and Indic languages. Voxtral Mini Transcribe 3 is a 4B-parameter model for fast batch transcription, and Voxtral Mini Transcribe Realtime 3 is built for agentic use with sub-500ms latency in over 35 languages. Both Small and Mini support context biasing, which lets developers supply key terms or free-form text to improve recall on domain-specific vocabulary without fine-tuning. They also handle diarization, identifying who spoke when, and end-of-turn detection that combines acoustic and semantic cues to decide when a speaker has finished.
Mistral claims Small Transcribe 3 is state-of-the-art on diarization and multilingual transcription, outperforming Whisper, Scribe v2 and Gemini 3.5 Transcribe. The published benchmarks use macro-averaged word error rate over the 45 languages shared by all seven compared models on FLEURS, with character error rate substituted for Japanese, Korean, Thai and Chinese. Diarization is measured with concatenated minimum permutation WER across AMI-IHM, AMI-SDM, Switchboard, Switchboard CallHome and AISHELL-4, using 30-minute versions of AMI and AISHELL for API compatibility. Keyphrase recall is evaluated on Earnings-22 with up to 100 distractor terms in the context bias. An Artificial Analysis WER Index aggregates VoxPopuli, Earnings-22 and AgentTalk. These are vendor-reported results, so independent replication would be needed before treating the rankings as settled.
On the speech generation side, Voxtral Mini TTS 3 is available in batch and realtime modes and supports 35+ languages and 50+ dialects. It adds voice cloning from 30+ minutes of sampled audio, custom prompt instruction following, contextual TTS and a vocabulary library. Explicit instruction variables cover emotion, pacing and emphasis. For cases where basic cloning is insufficient, Pro Voice adaptation aims to reproduce a speaker's character, emotional range and speaking style across languages from 30 minutes of audio. Voice presets are offered in 18 languages, including Emirati and Modern Standard Arabic, Brazilian Portuguese, Catalan, Mandarin, Dutch, American and British English, French, German, Greek, Hindi, Italian, Korean, Norwegian, Russian, Spanish and Vietnamese.
Mistral positions TTS pricing at $13 per million characters, which it says is cheaper than comparable offerings from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. The company published several quality-per-dollar charts covering enterprise failure states around numbers and unique names, pronunciation control for rare or invented words, multilingual quality across 11 languages and expressive English. Those charts rely on a Gemini-3.8-flash judge and vendor pricing pages, which makes them useful directional signals rather than neutral benchmarks.
The voice agent SDK is built on LiveKit and runs speech input, reasoning and speech output as a single streaming pipeline. Voxtral Mini Transcribe Realtime 3 listens and detects turn ends, a general model such as Mistral Small reasons and calls tools, and Voxtral TTS replies in the conversation's tone. Mistral says no stage waits for the previous one, so each turn completes in under one second, and the agent stops speaking when the caller interrupts. During a call the agent can verify identity, check a bill or book an appointment using function calling, web search and retrieval over internal systems. It handles inbound and outbound calls over SIP, runs in browsers and apps over WebRTC, and connects to existing telephony, IVR and contact-center software. Because each stage is a separate model, teams can swap the LLM, voice or language without rewriting the agent.
Pricing is listed per unit. Voxtral Mini Transcribe 3 batch costs $0.0035 per minute, the realtime variant $0.008 per minute and Small Transcribe 3 $0.0040 per minute. Voxtral Mini TTS 3 is $13 per million characters. The voice agent API costs $0.04 per minute with Mistral Small 4 for reasoning and $0.08 per minute with a larger model. On-premises or private cloud deployment is available by contacting Mistral. The models can also be tested in the Studio playgrounds for voice agents, transcription, speech generation and voice cloning.
A likely trade-off is that the modular pipeline gives flexibility but adds latency and cost relative to a single end-to-end speech model, and the per-minute agent pricing scales quickly for high-volume contact centers. The sovereignty story is a clear differentiator for European buyers: EU-hosted processing and self-deployment let organizations control where audio is processed, how data is handled and whether conversations are retained.
Why it matters: Teams building voice interfaces now get a single vendor stack for transcription, TTS and orchestration, with European hosting and on-prem options that may satisfy data-residency requirements. The low realtime transcription price and sub-second turn target make always-on agents more economically plausible, though the benchmark claims are vendor-reported and should be validated against your own audio. Developers should also compare the $0.04-$0.08 per-minute agent pricing with assembling the same pipeline from separate model providers.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Voxtral 3: Mistral's voice agent SDK for real-time speech in 50+ languages, accurate transcription, emotional TTS, deep tool use.