Mistral AI's third-generation speech suite: batch and real-time speech-to-text with diarization, expressive text-to-speech with voice cloning, and a LiveKit-based voice-agent SDK across 50+ languages.
Last verified · 2026-10-20 · by Moe Ameen
Voxtral 3 is the third generation of Mistral AI's Voxtral speech models, released on October 20, 2026. It is a family rather than a single model: Voxtral Small Transcribe 3 and the 4B-parameter Voxtral Mini Transcribe 3 handle batch transcription, Voxtral Mini Transcribe Realtime 3 streams transcription for live agents, and Voxtral Speech 3 is the text-to-speech model. Transcription covers 50+ languages, including all 24 official EU languages; the real-time model and the TTS model cover 35+ languages each.
What is new beyond accuracy is the surrounding capability. The transcribe models add speaker diarization (labeling who spoke when), context biasing (supplying key terms to lift recall without fine-tuning), and end-of-turn detection from acoustic and semantic cues. The TTS model shapes emotion, pacing, and emphasis from context and can clone a voice from sampled audio. A LiveKit-based voice-agent SDK ties speech-in, a reasoning model, and speech-out into one real-time pipeline with function calling, retrieval, and telephony over SIP or WebRTC. Mistral reports sub-500ms real-time transcription and sub-second agent turns, and transcription benchmarks ahead of Whisper, ElevenLabs Scribe v2, and Gemini 3.5 Transcribe — vendor-reported, so confirm on your own audio.
On pricing, transcription starts at $0.0035 per minute, TTS is $13 per million characters, and the full voice-agent API is $0.04–$0.08 per minute, through Mistral's EU, US, and global endpoints, with self-deployment possible for data-sovereignty needs. Be precise about the boundary: Voxtral 3 is a model and an SDK. It returns a transcript or an audio stream and stops — it writes no captions, cuts no clips, designs no carousel, drafts no blog, and publishes nowhere. Verify current specs and pricing against Mistral before relying on any figure.
The standout for content work is diarization. A raw transcript of a two-person interview or a four-person panel is a wall of text; a diarized Voxtral 3 transcript labels each speaker, which is exactly what you need to pull one guest's best line as a Quote Graphic or isolate a single answer as a clip. That is where [Kompozy](/) picks it up. Hand Kompozy the recording and it runs its own Whisper-based ASR to transcribe and caption, so you do not have to wire Voxtral yourself, then turns the strongest moments into captioned short clips, [Carousel](/glossary/hyperframes) posts, Quote Graphics, a [blog article](/ai-tools/koala-ai), and a newsletter — all governed by the [Persona Brief](/glossary/persona-brief) so the voice stays yours.
There is a second, complementary pairing on the audio side. Voxtral 3's expressive TTS and voice cloning are aimed at how a product speaks; Kompozy's own [Persona Shorts](/glossary/persona-shorts) and Persona HeyGen formats generate talking-head avatar video with native voice from an AI Influencer persona pool. If Voxtral 3 is how your app talks, Kompozy is how your brand shows up on camera — and both feed the same scheduled, multi-platform output queue. Voxtral 3 handles the speech; Kompozy handles finished, on-brand, published content.
Voxtral 3 is Mistral AI's family of speech models — batch and real-time speech-to-text plus a text-to-speech model — paired with a LiveKit-based voice-agent SDK that runs transcription, reasoning, and speech generation as one real-time streaming pipeline. It launched on October 20, 2026, with transcription across 50+ languages.
It is developer infrastructure for speech: transcribing audio (with diarization and context biasing), generating expressive text-to-speech with voice cloning, and building real-time voice agents and phone bots over the LiveKit SDK. It is a backend capability, not a finished app.
Not on its own. Voxtral 3 produces an accurate transcript but no clips, captions, carousels, or published posts. To turn that recording into captioned shorts, carousels, a blog, a newsletter, and scheduled posts across platforms, you use a content engine like Kompozy, which runs the transcription step for you with Whisper-based ASR.
By launch pricing, transcription is $0.0035/min (Voxtral Mini Transcribe 3), $0.0040/min (Voxtral Small Transcribe 3), and $0.008/min for the real-time model; TTS is $13 per million characters; and the voice-agent API is $0.04/min with Mistral Small 4 or $0.08/min with another reasoning model. Confirm current rates on Mistral's site.
Yes. Voxtral Mini Transcribe Realtime 3 streams transcription at sub-500ms latency across 35+ languages, and Mistral reports a full voice-agent turn completing in under one second, which is what makes live voice agents practical.