Mistral has introduced Voxtral 3, a software development kit designed to simplify the creation of real-time voice agents capable of multilingual interaction. The tool integrates speech transcription, natural language reasoning, and expressive speech synthesis into a single streaming pipeline, enabling developers to build systems that listen, interpret, act, and respond in live conversations.

The SDK supports more than 50 languages and 50 dialects, with transcription accuracy measured at 4.0% word error rate on the FLEURS benchmark for forced-language transcription. It includes two core models: Voxtral Small Transcribe 3 for high-precision speech-to-text conversion and Voxtral Mini TTS 3 for emotionally adaptive text-to-speech generation. Both models incorporate context-aware features such as speaker diarization, end-of-turn detection, and keyphrase biasing to improve performance in domain-specific environments.

A distinguishing feature is Pro Voice adaptation, which allows voice cloning from as little as 30 minutes of audio to replicate a speaker’s tonal qualities, emotional range, and linguistic style across languages. This capability aims to enhance personalization in customer service, virtual assistance, and hands-free computing scenarios.

Voxtral 3 is engineered for data sovereignty, offering deployment options through Mistral’s EU-hosted API or self-hosted infrastructure. Organizations can maintain control over audio processing locations, data retention policies, and integration with existing enterprise systems such as CRM platforms, scheduling tools, or internal databases via function calling.

The voice agent SDK operates as a unified pipeline, reducing latency by eliminating sequential processing delays. In testing, end-to-end response times fell below one second, with end-of-turn prediction combining acoustic and semantic cues to minimize misinterpretations during live interactions. Use cases include appointment scheduling, identity verification, and real-time support desk automation.

Pricing is usage-based: transcription starts at $0.0035 per minute for batch processing and $0.008 per minute for real-time streams. Speech generation costs $13 per million characters, undercutting several comparable commercial offerings. The SDK is now available for immediate testing via Mistral’s API console, with documentation and sandbox environments provided for developers.