Mistral AI has launched Voxtral 3, a speech model family and voice agent SDK that developers can use to build spoken conversations connected to business tools. Its transcription capabilities cover 50+ languages, while the SDK combines audio input, reasoning, tool calls and spoken responses in one streaming system. The models are available today through Mistral's API, hosted in the EU, or for deployment on an organization's own infrastructure.
Voxtral Mini Transcribe 3 has 4 billion parameters and costs $0.0035 per minute for transcription. Language coverage differs across the family: Voxtral Mini TTS 3 generates speech in 35+ languages and 50+ dialects, while Voxtral Realtime Transcribe 3 handles transcription in over 35 languages. Mistral says the realtime model delivers transcriptions with latency below 500 milliseconds, including when identifying speakers and using supplied vocabulary.
The voice agent SDK is intended for applications such as service desks and virtual assistants, with live conversations supported over the web or phone. Mistral says developers can connect agents to their own information and tools so they can look up answers, work with business systems and carry out tasks involving multiple steps during a conversation. Organizations can also control where conversations are processed and whether audio is retained.
For transcription, both the Small and Mini models accept key terms and additional text to help recognize specialized vocabulary without fine-tuning. Voxtral Small Transcribe 3 also identifies speakers and detects when a person has finished talking, which Mistral says helps agents respond promptly. The expanded transcription coverage includes 24 official EU languages as well as Asian and Indic languages with fewer resources.
On the speech generation side, Mistral says Voxtral Mini TTS 3 can change a voice's emotional delivery, speed and emphasis to suit the conversation. The Voxtral 3 family's Pro Voice adaptation uses 30 minutes of recorded audio to reproduce a speaker's vocal character, emotional expression and style across languages, according to the company.
Mistral's published multilingual test gives Voxtral Small Transcribe 3 a 4.1% error rate when the language is specified and 4.0% when detected automatically. Those results cover 45 languages shared by the seven models compared, with character error rates substituted for word error rates for Japanese, Korean, Thai and Chinese.
