Mistral Launches Voxtral 3 Speech Model Suite and Voice Agents API

On October 21, Mistral released the Voxtral 3 speech model family along with a Voice Agents API, covering speech recognition, speech synthesis, and voice agent building in one go. The transcription and TTS components are claimed to reach SOTA on multiple benchmarks, and a high-fidelity voice cloning capability has been added—marking Mistral's full push from pure-text/multimodal models into real-time voice interaction.

Confirmed

• The Voxtral 3 family includes Transcribe (speech recognition) and Speech (speech generation) models, accompanied by the new Voice Agents API (m1)
• A voice Agent SDK built on LiveKit is also available: Transcribe handles listening, a reasoning model thinks and calls tools, and Speech generates the reply, with all three streaming together into one pipeline for natural conversation (m2)
• Voxtral Small Transcribe 3 achieves SOTA in multilingual transcription, speaker diarization, and contextual biasing, and can detect when a speaker finishes talking, enabling voice agents to respond without delay (m3)
• Voxtral Mini TTS 3 delivers SOTA-level expressiveness across 35+ languages and 50+ dialects; it supports implicit instruction following (reading emotion, pacing, and emphasis from the conversation) as well as explicit tags like [emotion=sad][pace=slow] for line-by-line control (m4)
• Pro Voice voice adaptation: record up to 30 minutes of audio to replicate the user's voice with very high fidelity across all supported languages (m5)

Why it matters

• Voxtral 3 chains transcription, reasoning, and speech generation into a streaming pipeline, directly targeting the real-time voice APIs of vendors like OpenAI and Google, and giving developers an out-of-the-box voice agent solution
• End-of-speech detection plus coverage of 35+ languages and 50+ dialects shows Mistral is clearly aiming at the multilingual real-time voice assistant market, while Pro Voice voice cloning fills in a key capability for personalized voice interaction
• All of the above comes from Mistral's official announcement; the performance claims (SOTA) still await third-party evaluation

2026-10-21 ~ 2026-10-21 · 9 related posts

• [source] Mistral launches Voxtral 3 speech models and new Voice Agents API — MistralAI · 2026-10-21
• Voxtral Small Transcribe 3 claims SOTA in multilingual transcription and diarization — MistralAI · 2026-10-21
• Voxtral Mini TTS 3 hits SOTA expressiveness in 35+ languages with tag-based control — MistralAI · 2026-10-21
• Mistral's Pro Voice clones your voice from 30 minutes of audio — MistralAI · 2026-10-21
• Voxtral 3 ships a LiveKit-based voice agent SDK streaming listen-think-speak — MistralAI · 2026-10-21
• [source] Mistral's Voxtral 3 speech models: #3 STT accuracy, fastest TTS at $13/1M chars — ArtificialAnlys · 2026-10-21
• Mistral releases Voxtral 3: voice agent SDK and speech models in 50+ languages — arthurmensch · 2026-10-21
• [source] Mistral releases full suite of state-of-the-art voice models, betting voice is AI's key interface — arthurmensch · 2026-10-21

1 near-duplicate retellings: GuillaumeLample