Mistral Launches Voxtral 3, a Complete Toolkit for Voice Assistants That Talk Back
Mistral's new Voxtral 3 family bundles speech recognition, speech output and an SDK for real-time voice agents, with EU hosting or your own servers as options.
Mistral AI has released Voxtral 3, a family of models built for voice assistants that can listen, think and answer out loud in real time. According to The Decoder’s summary of the announcement on October 20, it has three parts: Voxtral Small Transcribe 3 turns speech into text, Voxtral Mini TTS 3 turns text back into speech, and a software kit (an “SDK”) built on LiveKit ties them together into one live pipeline.
The numbers on offer are modest but clear. Transcription covers more than 50 languages, while the speech output supports over 35 languages and 50 dialects and can adapt its tone to the emotion of a conversation. Transcription pricing starts at $0.0035 per minute, which works out to roughly twenty cents per hour of audio. Because the pieces run as one streaming pipeline, an agent can also call outside systems mid-conversation, for example to book an appointment or look up an invoice, instead of waiting until you stop talking.
The part that stands out for European readers is where it can run. Companies can use EU-hosted APIs or run the whole stack on their own infrastructure, which matters for anyone who is not comfortable sending recorded customer calls to a server abroad. Mistral is a French company, so this fits its usual pitch.
Some gaps are worth knowing about. The coverage names no benchmark scores, no response-time figures and no model sizes. It is also unclear whether the $0.0035 price covers just transcription or the whole stack, and licence terms for self-hosting were not spelled out in the summary we could check. Treat the claims as an announcement, not a test result.
What’s behind it: a good voice assistant is harder than it sounds. It must hear you accurately, decide what to do, call the right tool and answer fast enough that the pause does not feel awkward. Doing that with three separate products from three vendors is fiddly. Mistral is betting that a single bundle is easier to build with. In plain terms, this is less a new brain and more a ready-made set of ears, a mouth and the wiring between them.
What this means for you: if you simply use chatbots, nothing changes today. If you build tools for a small business, such as a phone line that answers opening-hours questions or books appointments, Voxtral 3 is worth a look, especially if keeping data in the EU is a requirement. Try it on a handful of your own recordings first, with accents and background noise, and compare accuracy and delay before you commit.
Sources
OpenAI Says GPT-6 Has Got Much Better at COBOL and Other Old Programming Languages
OpenAI reports big gains on rarely used languages like COBOL and Fortran, including a jump on a COBOL migration test. All figures are the company's own.
