Skip to content
NewsModel release

Mistral ships Voxtral 3, a self-hostable voice-agent pipeline with 30-minute voice cloning

· by Pondero Newsdesk

The short version

Mistral released Voxtral 3, bundling speech transcription, reasoning, and speech synthesis into one streaming pipeline, plus a Pro Voice feature that clones a speaker from 30 minutes of audio.

Mistral ships Voxtral 3, a self-hostable voice-agent pipeline with 30-minute voice cloning

Mistral released Voxtral 3, a family of speech models and a voice-agent SDK that runs transcription, reasoning, tool calling, and speech synthesis as one real-time streaming pipeline, per Mistral's announcement. Most teams building voice agents today still stitch those jobs together from separate vendors.

What

Voxtral Small Transcribe 3 covers speech-to-text in 50-plus languages, including 24 official EU languages, per Mistral. It adds speaker diarization and context biasing for domain vocabulary. A separate end-of-turn signal tells an agent when a caller has actually stopped talking, not just paused for breath. On Mistral's own FLEURS benchmark, the model beat Whisper large-v3 and Scribe v2 on word-error rate. Gemini 3.5 Transcribe trailed too, in a comparison the company ran and published itself. A smaller, 4B-parameter version handles the same transcription job at $0.0035 per minute. Voice comes out the other end through Voxtral Mini TTS 3.0, which spans 35-plus languages and 50-plus dialects at $13 per million characters. It sets emotion and pacing from conversational context automatically. Emphasis follows the same rule, rather than a developer scripting it line by line. Pro Voice, a separate adaptation feature, goes further: give it 30 minutes of sampled audio and it reproduces a speaker's character and emotional range. Speaking style carries across languages too. Everything ships today through Mistral's EU-hosted API or as a self-deployed install on a customer's own servers.

Self-hosting and price are the pitch, not raw accuracy

For teams evaluating a voice-agent stack, the self-deployment option is the real change: it lets a hospital, bank, or government contractor keep call audio off a third-party server entirely, something the EU-hosted API alone does not guarantee under every compliance regime. On price, Mistral cited a third-party comparison from BeSimple AI's Voice Code Bench showing Voxtral TTS costing less per character than OpenAI's gpt-4o-mini-tts and Deepgram Flux at comparable accuracy. Eleven v4 and several Gemini TTS tiers priced higher still, per the benchmark Mistral published. That combination, EU data residency plus a lower per-character rate, is the argument Mistral is making to enterprises already running voice agents on ElevenLabs or OpenAI's realtime stack, not a claim of outright superior voice quality. Pro Voice cuts the other way: a feature that reproduces a real person's speaking style from half an hour of audio is also the feature most likely to draw consent and impersonation scrutiny once it reaches wider use, something Mistral's announcement does not address.

What to watch next

Watch for independent latency and quality benchmarks against ElevenLabs and OpenAI's Realtime voice stack, since the comparisons published so far are Mistral's own or ones Mistral chose to cite. Also watch whether Pro Voice draws consent or misuse scrutiny once third parties start testing it on real speakers.

Sources

  • Voxtral 3: Mistral's product announcement, primary