Transcription, speech synthesis and conversational agents in one set
France's Mistral AI on Oct. 20 (Oct. 21 Japan time) announced "Voxtral 3," a family of speech models. The four models are "Small Transcribe 3" and "Mini Transcribe 3" for batch transcription, "Mini Transcribe Realtime 3" for transcribing speech as it is spoken, and "Mini TTS 3" for turning text into speech, all launched as a trial offering.
The company also released an API for building agents that converse by voice. It runs the whole flow in a single pipeline: transcribing speech, having a conversational model decide on the answer and how to use tools, and replying by voice. Mistral said each exchange completes in under a second and that the agent stops speaking if interrupted. It can connect to phone lines as well as browsers and apps, but phone access is initially limited to business customers.
Names in the API list 2.5 hours before the announcement
The names of the four models leaked out before the announcement. At around 12:30 a.m. Oct. 21 Japan time, @aitrackerbot, an X account that tracks the appearance of models, reported that four names had been listed in Mistral's API model list. The official announcement came about two and a half hours later.
Transcription error rate less than a third of Whisper's
In Mistral's measurements, the word error rate when transcribing speech in 45 languages was 4.0% for Small Transcribe 3 and 4.5% for Mini Transcribe 3. That is well below the 14.9% of "Whisper large-v3," released by OpenAI, and on par with Microsoft's "MAI-Transcribe 2." The company also improved accuracy in distinguishing multiple speakers, with Mini's speaker confusion rate falling from 30.9% in the previous generation to 18.6%.
Transcription from 0.35 cents a minute
Transcription costs $0.0035 to $0.008 per minute, and speech synthesis costs $13 per 1 million characters. The conversational agent costs $0.04 per minute with the standard conversational model "Mistral Small 4" and $0.08 with a larger model. Mistral said its speech synthesis pricing is lower than comparable models from ElevenLabs, OpenAI and Google.
| Type | Price |
|---|---|
| Mini Transcribe 3 | $0.0035 per minute |
| Small Transcribe 3 | $0.004 per minute |
| Mini Transcribe Realtime 3 | $0.008 per minute |
| Mini TTS 3 | $13 per 1 million characters |
| Conversational agent | $0.04 to $0.08 per minute |
Weights not released; performance measured in-house
The previous generation of Voxtral had its weights released, but Voxtral 3 does not. Companies that want to run it on their own infrastructure need to contact sales. All performance figures were measured by Mistral itself, and word-level timestamp output, available in the previous generation, has been dropped.
Speech models arrived on the same day. ElevenLabs also released "Eleven Realtime v1," a new model for voice agents, that day.
