The startup announced a new suite of transcription and generation models called Voxtral 3, alongside a streaming SDK designed to deliver sub‑second turn completion for voice assistants.

Mistral AI has unveiled Voxtral 3, a new suite of speech models that promises higher accuracy in transcription and more natural voice generation, alongside a low‑latency streaming SDK aimed at powering next‑generation voice assistants.

Voxtral 3: Enhanced Speech Capabilities

The Voxtral 3 family includes both large and compact models for automatic speech recognition (ASR) and text‑to‑speech (TTS). According to the company, the ASR models reduce word error rates by a few points compared with the previous generation, while the TTS models deliver smoother prosody and clearer articulation across multiple languages.

Sub‑second Turn Completion SDK

Mistral’s new SDK streams audio to the model and returns partial hypotheses in real time, enabling voice agents to respond within sub‑second intervals. The design focuses on minimizing buffering and leveraging efficient model quantisation to keep compute costs low.

Developers can integrate the SDK via a simple API that supports both on‑device and cloud‑based deployments, giving flexibility for edge‑focused products or scalable server‑side solutions.

Potential Use Cases

  • Voice‑controlled smart home devices
  • In‑car infotainment systems
  • Customer‑service chatbots with spoken interaction
  • Accessibility tools for real‑time captioning

Mistral AI positions Voxtral 3 as a drop‑in upgrade for existing voice applications, promising smoother user experiences without the need for extensive retraining.

For readers interested in Mistral’s broader AI roadmap, see the company’s profile on Mistral AI on Momentum and explore related stories in the AI category.

e-acc.ai daily coverage of Mistral AI’s Voxtral 3 launch