Mistral AI announced Voxtral 3 on October 20, 2026. The release includes speech-to-text and text-to-speech models plus a voice agent SDK built for real-time, multi-turn conversations in more than 50 languages.
Voxtral 3 Components and Technical Specifications
The stack comprises Voxtral Transcribe 3 in Small and Mini variants for batch and realtime use, Voxtral Speech 3 for synthesis, and a LiveKit-based SDK. The SDK connects listening, reasoning with tool-calling via Mistral Small, and speaking in a single pipeline. Realtime transcription targets sub-500 ms latency in 35 languages, while the full pipeline aims for sub-second turn times.
Transcription supports 50 languages, including all 24 official EU languages and several low-resource Asian and Indic languages. Features include speaker diarization, context biasing, and end-of-turn detection. Text-to-speech covers 35 languages and 50 dialects with emotional control and both implicit and explicit instruction following. Pro Voice allows high-fidelity cloning from up to 30 minutes of reference audio.
API pricing lists Mini Transcribe 3 at approximately $0.0035 per minute for batch and $0.008 per minute for realtime. Small Transcribe 3 is priced at $0.004 per minute. TTS and Speech 3 cost $13 per million characters for both batch and realtime. The models are available through Mistral’s EU-hosted endpoints and for self-deployment on customer infrastructure.
Performance Claims and Enterprise Applications
Mistral states that Voxtral 3 achieves leading results against Whisper large-v3, OpenAI, ElevenLabs, Deepgram, and Gemini models on accuracy, diarization, expressiveness, and multilingual benchmarks. The company cites internal and external evaluations published on its announcement page.
The models target contact centers, virtual assistants, and workflow automation where low latency and data sovereignty matter. The EU-hosted option supports compliance needs for regulated industries. Self-deployment allows organizations to keep voice data inside their own infrastructure.
Access is immediate through Mistral’s API, the company’s Studio playground, and platforms such as OpenRouter. Developers can integrate the SDK to build agents that handle tasks beyond simple question answering by calling external tools in real time.
Market Context and Next Steps for Users
Voice agents have moved from experimental pilots to production systems in customer service and internal operations. Organizations now evaluate latency, language coverage, and deployment flexibility when selecting providers. Voxtral 3 enters this market with explicit emphasis on European data residency and competitive pricing for high-volume transcription.
Teams evaluating the release should test latency and diarization on their own audio samples. The Mini realtime variant offers the lowest entry cost for initial pilots, while the Small model provides higher accuracy for production workloads. Documentation and playground access are live on Mistral’s site as of the announcement date.
Frequently asked questions
What languages does Voxtral 3 support for transcription?
Voxtral Transcribe 3 covers more than 50 languages, including all 24 official EU languages plus selected low-resource Asian and Indic languages, with diarization and context biasing features.
How much does the Voxtral 3 API cost?
Mini Transcribe 3 is priced at roughly $0.0035 per minute in batch mode and $0.008 per minute in realtime. TTS and Speech 3 are charged at $13 per million characters.
Can Voxtral 3 run on customer infrastructure?
Yes. The models are offered for self-deployment in addition to Mistral’s EU-hosted API endpoints, supporting data sovereignty requirements.
What latency does the realtime transcription target?
Mini Transcribe Realtime 3 aims for sub-500 ms latency across 35 languages, with the overall voice agent pipeline designed for sub-second conversational turns.
By YasTech Editorial
Comments 5