Voxtral 3 for video people: cheaper captions that know your product names, and narration with emotion tags

Mistral launched Voxtral 3 on Oct 20 (the launch post passed 2,000 likes within hours, and @RoundtableSpace picked it up). The headline is a voice-agent SDK for phone and web assistants. We care about the two models underneath it, because transcription and voiceover are where a lot of agent video pipelines spend their money and their mistakes.

Transcription first. Voxtral Small Transcribe 3 and the 4B Mini Transcribe 3 cover 50+ languages with diarization (who spoke when) and context biasing: you pass key terms and free text so it spells your product, your guests' names and your jargon correctly. Mistral's own charts put both at around 4% word error rate on a 45-language FLEURS average, against about 16% for Whisper large-v3. On their keyphrase test, biasing lifts recall from 81% to 95%. For captions and transcript-driven cutting, that's the error that usually costs you a manual pass. Prices: $0.0035 per minute (Mini, batch) and $0.004 per minute (Small). There's also a realtime model at $0.008 per minute.

Then voiceover. Voxtral Mini TTS 3 costs $13 per million characters. Our rough maths puts a minute of narration at about a cent. The model page lists 34 emotion tags, 2 pace tags, streaming with about 90 ms to first audio, and zero-shot or 'professional' cloning (the docs recommend 30 minutes to 3 hours of clean audio for the latter). Saved cloned voices are kept 30 days by default. One inconsistency to check before you plan multilingual dubs: the blog says 35+ languages, while the model page says 9.

Caveats. All of these numbers are Mistral's own benchmarks. Voxtral 3 is API-only for now (self-hosting means talking to their sales team), and the earlier open Voxtral-4B-TTS checkpoint can't clone voices because its encoder weights were never released. If you need local dubbing, our VoiceStudio post covers that route. Only clone voices you have consent to use. Spotted, not tested. Credit Mistral AI.