Technical
Mistral's Voxtral 3: Speech-to-Text, Text-to-Speech, and a Voice-Agent SDK in One Release
ยท RenderBob team
Mistral released Voxtral 3 on 20 October, covering transcription, text-to-speech, and an SDK that chains listening, reasoning, and speaking. Test the transcripts first.
Mistral released Voxtral 3 on 20 October, a family covering transcription, text-to-speech, and an SDK that chains listening, reasoning, and speaking. For post-production, the transcription side is the part to test first. The figures below are from Mistral's news page.
What the page lists
- Voxtral Small Transcribe 3, the higher-accuracy speech-to-text model, at $0.0040 per minute.
- Voxtral Mini Transcribe 3, a 4-billion-parameter model for batch transcription, at $0.0035 per minute.
- Voxtral Mini Transcribe Realtime 3, streaming, at $0.008 per minute, with sub-500 ms latency.
- Voxtral Speech 3, text-to-speech in batch and real time, at $13 per million characters. The page also calls this model Voxtral Mini TTS 3.
- A voice agent SDK, built on LiveKit, chaining listening, reasoning, tool calling, and speaking. Mistral prices it at $0.04 per minute with Mistral Small 4, or $0.08 per minute with another reasoning model, and says each turn completes in under a second.
Transcription is listed at 50-plus languages. The same page gives two counts for realtime transcription: the body says over 35 languages, and a model card says 40-plus. Text-to-speech is listed at 35-plus languages and 50-plus dialects. Mistral says Small Transcribe 3 outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe on multilingual transcription and diarisation. The page includes an Artificial Analysis word-error chart, and several of the speech charts are marked as Mistral's own evaluations.
What the page does not say
It does not state a licence, or whether these weights are open. It says you can use the API or deploy on your own infrastructure, and that a full on-premises or private-cloud stack means contacting Mistral. Voice cloning is described from 30-plus minutes of sampled audio, and a Pro Voice option from 30 minutes. The page describes no consent check or verification step for cloning.
For film and video work
Transcripts drive editing. Our posts on the transcript layer and on suggestive editing rely on accurate, speaker-labelled transcripts. Diarisation quality is worth testing on your own interviews and on non-English material. The per-minute prices are low enough that batch transcription of rushes is practical to try. Cloning without a described consent step is a risk for any production using real voices. Put your own written consent process in place first.
A quick test
Take 20 minutes of messy multi-speaker audio with crosstalk and two accents. Run it through Small Transcribe 3 and your current tool, then count speaker-attribution errors and name errors. When those transcripts feed a picture cut, RenderBob lets you choose local or cloud for each clip and keep the references with the studio.
What to watch
A licence statement, a published independent benchmark, and any position from Mistral on voice-cloning consent.
RenderBob 3 and this story
RenderBob 3 holds the cut and its records together. Whichever speech tool you choose, log its terms beside the prompts kept on your generated shots.
More from the blog
- FLUX 3 Video Keyframes: How to Pin Frames, Storyboard Beats, and What the Docs Say Will Drift
Runware's FLUX 3 Video docs let you pin up to 10 images by frame index or by timestamp. The page calls that strong guidance, and it marks storyboards and morphs experimental.
- Fish Audio Drama 3: A 35B Open-Weight Voice Model You Direct in Plain Language
Fish Audio announced Drama 3 on 13 October. You write performance direction in square brackets, and one text can carry up to 16 speakers. The weights are on Hugging Face.