Voxtral 3: $0.0035/min transcription, sub-500ms realtime STT, no stated licence

Mistral frames the problem as getting four things to work together reliably in real time: listening across languages, accents, speakers and interruptions; reasoning about intent; speaking with emotion; and taking multi-step action through connected systems. The target uses it names are contact-center agents, service desks, virtual assistants and hands-free copilots. You can call the models through Mistral's EU-hosted API or deploy them on your own infrastructure. Mistral says both options give control over where audio is processed, how data is handled and whether conversations are retained.

Mistral says Voxtral Small Transcribe 3 achieves state-of-the-art results in diarization and multilingual transcription, including English, French and Spanish. It says the model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. The announcement does not say which Whisper variant it compared against. The evidence takes the form of four charts:

The announcement text does not state the WER, cpWER or recall values, so the margin over Whisper cannot be checked from the text. The forced versus inferred split matters in production. If your traffic arrives without a language tag, the inferred column is the one to read.

Context biasing lets you pass key terms and free-form text to improve recall on technical terms without fine-tuning. Commenters on Hacker News welcomed native biasing. One said that at enough volume they would still self-host and do their own biasing. Another said earlier Voxtral models did not hallucinate text when there was no speech. The announcement does not say whether that still holds.

The only latency figure in the announcement belongs to Voxtral Realtime Transcribe 3: sub-500ms in over 35 languages. The announcement does not say how that latency was measured. Its accuracy is charted as macro-average WER on FLEURS 35, with Voxtral Mini Transcribe 3 included as an offline baseline. End-of-turn prediction is built in and uses both acoustic and semantic information. Mistral says this allows low-latency agents while minimizing false cut-offs, but it gives no false cut-off rate.

The announcement gives no figure for the full voice agent loop: transcribe, reason, call tools, then synthesize. One commenter asked how real-time the demo conversation actually is, since it shows no lag between responses. Another argued that the real value is the integration of a real-time reasoning model and whatever hides tool-call latency.

The announcement text does not name the reasoning model. One commenter referred to it as Voxtral Large; the announcement does not confirm this. That commenter also asked whether the model emits a special token or runs a command when it needs help, and the source does not say. If you are budgeting a latency target, you will have to measure the full loop yourself.

Mistral says the TTS model costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. It also claims state-of-the-art expressiveness and instruction following. The model infers emotion and voice style from the text. Explicit instruction variables cover emotion, pacing and emphasis, and the demos use non-standard phoneme hints to control how rare or made-up names are pronounced. Other listed features are voice cloning from 30+ minutes of sampled audio, contextual TTS and a vocabulary library.

Four quality-per-dollar charts support the claims. Voice Code Bench, made by BeSimple AI, uses 300 English prompts focused on customer support failures with numbers and unique names. The other three are internal evals: pronunciation control (300 sentences, each with one rare or made-up word), multilingual (377 prompts across 11 languages) and expressive English (237 prompts). All four use gemini-3.8-flash as the judge, so the scores reflect one LLM's judgment rather than human listener ratings.

The demo comparisons target number and date reading. For a flight prompt, the Voxtral 1 output shown reads '10 over 10' and '21,000pm'. The Voxtral 3 output reads 'October 10th' but still renders the times as '21pm' and '22pm'. One commenter disagreed with the direction of the change, saying version 1 sounded like actual people and the new voices sound generic.

The announcement offers self-deployment, but it names no licence and does not mention downloadable weights. Commenters on Hacker News described the models as closed, and one said this was unlike the previous version. That reading comes from commenters. The source does not confirm it.

Commenters who need local inference listed alternatives, including Qwen ASR, QwenTTS, Kokoro and kyutai's pocket-tts. The commenter who suggested pocket-tts noted it covers only a few languages. Another commenter argued that TTS models carry a different threat model from LLMs, which makes open release harder. One commenter said that for large European companies that want to control their data, Voxtral 3 may be one of the best options now. The announcement's own offer is EU hosting and the ability to keep audio inside your environment, and it does not describe commercial terms for self-deployment.

If your Whisper-plus-LLM stack mostly struggles with diarization, domain vocabulary or turn-taking, Voxtral Small Transcribe 3 and Voxtral Realtime Transcribe 3 target those problems directly, based on the stated features. If your stack depends on open weights, it is unclear whether the release qualifies until Mistral publishes licence terms.