ميسترال تطلق Voxtral 3 لبناء وكلاء صوتيين، ومن أصواته الجاهزة العربية الإماراتية والفصحىMistral launches Voxtral 3 for building voice agents, with ready-made Emirati and Modern Standard Arabic voices

Mistral has launched Voxtral 3, a family of speech-to-text and text-to-speech models, together with a developer kit (SDK) for building voice agents that listen, think, carry out tasks and reply out loud, all in one conversation.

The transcription model supports more than 50 languages, tells who said what in a recording and knows when a speaker has finished. Its small Mini version has 4 billion parameters and costs $0.0035 a minute. The speech model supports more than 35 languages and 50 dialects, with voices that change emotion, pace and emphasis to fit the conversation, at $13 per million characters.

For us, ready-made voices come in 18 languages, including Arabic in two versions: Emirati and Modern Standard Arabic. Mistral says a voice agent can verify a caller's identity, check a bill or book an appointment, that each turn of the conversation finishes in under a second, and that the agent stops talking when the caller cuts in.

The models are available today through Mistral's API, or organisations can run them on their own servers. But the performance and price comparisons with rivals are Mistral's own figures, and no Saudi dialect is listed among the ready-made voices it named.