Mistral AI launches Voxtral 3 voice agent stack with speech-to-text, text-to-speech and a LiveKit-based SDK
Mistral released Voxtral 3, a family of voice models: Voxtral Small Transcribe 3 (speech-to-text, 50+ languages), Voxtral Mini TTS 3 (text-to-speech, 35+ languages and 50 dialects with emotional adaptation), and a voice agent SDK based on LiveKit. These form a streaming pipeline covering listening, reasoning, tool use and speaking. It can be self-hosted or used via EU-hosted APIs. Transcription pricing starts at $0.0035 per minute.
Entities: Mistral AI, Voxtral 3, Voxtral Small Transcribe 3, Voxtral Mini TTS 3, LiveKit, The Decoder
0 primary
What happened
Mistral AI released Voxtral 3, a set of voice models. It has three parts: Voxtral Small Transcribe 3 (speech-to-text, 50+ languages), Voxtral Mini TTS 3 (text-to-speech, 35+ languages and 50 dialects, with emotional adaptation) and a voice agent SDK built on LiveKit. Together they form a streaming pipeline for listening, reasoning, calling tools and speaking. Customers can self-host or use EU-hosted APIs. The only price given is transcription from $0.0035 per minute. This account comes from The Decoder, not Mistral's own announcement, and no primary links were captured.
Why it matters
Developers building voice agents, especially European enterprises with data residency or privacy requirements, get one vendor offering transcription, speech and an agent framework, with a self-hosting option. The published transcription price makes a basic cost estimate possible. Text-to-speech and agent pricing are not stated, so total cost per call cannot yet be worked out. The impact is moderate: this is the third version of an existing model family in a crowded market, and it is not a clear step change.
What is noise
"Complete real-time voice agent stack" is Mistral's framing. The SDK is built on LiveKit, so the orchestration layer is largely existing open tooling. Language and dialect counts say nothing about accuracy, latency or how natural the voices sound, and "emotional adaptation" is an unverified marketing claim. No benchmarks or independent tests appear in the coverage.
Watch next
- 01Independent benchmarks against ElevenLabs, OpenAI, Deepgram and Google: word error rate across languages, end-to-end latency and voice quality, within the next few weeks.
- 02Full pricing for text-to-speech and the agent runtime, plus the self-hosting licence terms and hardware needs, which decide whether the data-control claim is practical.
- 03Named production customers or developer adoption (SDK repository activity, LiveKit integrations) over the next one to two quarters.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680
- OpenAI launches GPT-6 Sol and Luna at half the token price of GPT-5.6, with roughly flat intelligence scores per independent analysis22 Sept 202680