Mistral Launches Voxtral 3 for Real-Time Multilingual Voice Agents
Mistral AI has introduced Voxtral 3, a new family of speech-to-text models, text-to-speech generators, and a voice agent SDK built for real-time multilingual conversations across 50+ languages.
Mistral AI has officially released Voxtral 3, a comprehensive suite of advanced speech-to-text models, text-to-speech generators, and a dedicated voice agent software development kit. Designed to facilitate multi-step, real-time conversations, the new lineup supports more than 50 languages and aims to bridge the gap between accurate audio transcription, deep contextual reasoning, and expressive voice synthesis.
Building dependable voice agents that can listen across various accents, reason through complex requests, and execute actions across connected systems has traditionally been a formidable engineering hurdle. Voxtral 3 attempts to solve this challenge by integrating listening, reasoning, tool calling, and speaking into a unified streaming pipeline built on LiveKit. Organizations can run the entire stack using Mistral's EU-hosted API or deploy it directly on their own infrastructure for complete data sovereignty.
Core Components of the Voxtral 3 Family
The speech-to-text lineup features models like Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3, which deliver advanced multilingual transcription and speaker diarization—identifying who spoke and when. These models incorporate context-biasing capabilities that allow developers to supply domain-specific key terms and free-form text, significantly improving recall on technical vocabulary and proper nouns without requiring custom fine-tuning. Additionally, Voxtral Realtime Transcribe 3 achieves sub-500ms latency while utilizing acoustic and semantic cues for native end-of-turn detection.
On the generation side, Voxtral Mini TTS 3 provides expressive text-to-speech capabilities across more than 35 languages and 50 dialects. The model implicitly interprets semantic context to adapt its emotion, pacing, and emphasis without needing explicit prompt instructions. For professional use cases, Voxtral Pro Voice replication can reproduce a specific speaker's distinctive style, character, and emotional range across languages using just 30 minutes of audio samples.
What it means for developers
Developers building customer service contact centers, automated service desks, or hands-free virtual assistants can leverage the Voxtral voice agent SDK to orchestrate live conversations over web or telephony channels. Because the architecture treats listening, reasoning, and speaking as modular stages, engineering teams can swap underlying language models, alter voices, or change languages without rewriting their core application logic. Developers can try top AI models cheaply through one API at https://apixoai.online.
Pricing is structured per usage across batch and real-time tiers. Voxtral Mini Transcribe 3 is priced at $0.0035 per minute for batch workloads, while the real-time transcription variant costs $0.008 per minute. Voxtral Speech 3 is available for $13 per one million characters for both batch and real-time usage. Developers can access the models via Mistral's API endpoints or choose self-deployment options for high-security environments.
Source: Voxtral 3 — Mistral AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
