Voxtral 3

Mistral发布了Voxtral 3，一套集成语音识别、自然语言推理、工具调用和语音生成功能的新型语音AI模型与语音智能体SDK1。该产品支持50多种语言进行实时多步骤对话，涵盖24种官方欧盟语言以及多种亚洲和印度语言1，可应用于客服、虚拟助手等场景1。用户可选择通过Mistral的欧洲托管API或自部署方式运行该系统，以确保数据主权1。

产品包含多个专用模型，定价各异。语音转文本模型Voxtral Small Transcribe 3的价格为每分钟$0.004，Mini版本为$0.0035/分钟1；文本转语音模型Voxtral Mini TTS 3支持35+种语言和50+种方言，价格为每百万字符$131。实时转录模型Voxtral Mini Transcribe Realtime 3可实现低于500毫秒的延迟，定价为每分钟$0.0081。此外，该系统支持Pro Voice适配功能，仅需30分钟的音频样本即可复制使用者的语音特征1。

Mistral has unveiled Voxtral 3, a comprehensive suite of voice artificial intelligence models and an intelligent agent software development kit designed for real-time, multi-step conversations across more than 50 languages 1. The platform integrates speech recognition, natural language reasoning, tool invocation, and speech synthesis capabilities, making it suitable for applications such as customer service and virtual assistants 1.

The Voxtral 3 suite comprises several specialized models with distinct pricing structures 1. The Voxtral Small Transcribe 3 model, which converts speech to text, is priced at $0.004 per minute, with a Mini variant available at $0.0035 per minute 1. For text-to-speech conversion, the Voxtral Mini TTS 3 model supports 35 or more languages and 50 or more dialects at a rate of $13 per million characters 1. Additionally, a real-time transcription model, Voxtral Mini Transcribe Realtime 3, delivers latency below 500 milliseconds at $0.008 per minute 1. The platform includes Pro Voice adaptation functionality, which requires only 30 minutes of audio to capture and replicate a speaker's vocal characteristics 1.

Deployment flexibility is central to Voxtral 3's design, as the models can operate either through Mistral's European-hosted API or through self-deployed infrastructure, allowing organizations to maintain data sovereignty 1. The language support extends to 24 official European Union languages alongside multiple Asian and Indian language variants 1.