Mistral AI Releases Voxtral 3: Speech-to-Text, TTS and a Voice Agent SDK
MarkTechPost Asif Razzaq ● Covered by 2 sources
Mistral launched Voxtral 3, a new speech and voice-agent bundle. It mixes transcription, TTS and calling tools, but realtime accuracy still trails some rivals.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral has pushed out Voxtral 3, a speech model family plus a voice agent SDK built on LiveKit. The package includes three transcription models, one text-to-speech model and a pipeline meant to handle listening, reasoning, tool calls and speaking in one flow. All four models are already in public preview on the Mistral API.
The lineup is split by job. Voxtral Small Transcribe 3 is the flagship batch model, with translation, speech Q&A, summarization and function calling in the docs. Voxtral Mini Transcribe 3 is the cheaper 4B option. Voxtral Mini Transcribe Realtime 3 handles streaming ASR over WebSocket with end-of-turn detection, while Voxtral Speech 3 is the TTS model, also referred to as Voxtral Mini TTS 3 in the post.
Mistral says all transcription models support diarization and context biasing, and batch transcription covers 50 languages, including 24 official EU languages. The company also says Mini is meant for resource-constrained environments. The models can run on Mistral’s EU, US and global endpoints, or on customer infrastructure through Mistral sales. GPU and VRAM requirements were not disclosed.
On its own benchmarks, Mistral puts Small near the top of speech-to-text rankings. It reports 2.1% AA-WER for Small Transcribe 3, with strong diarization and multilingual batch transcription. The model’s self-reported FLEURS 45 score is 4.1% WER, matching MAI-Transcribe-2 on one setting and beating Whisper large-v3. But the realtime side is less tidy: at 240 ms on FLEURS 35, the realtime model hits 7.4% WER, while Gemini 3.5 Realtime sits at 4.2%.
The voice agent SDK is the more ambitious part. Mistral says each turn completes in under 1 second, and the agent stops talking when a caller interrupts. It works over SIP for phone calls and WebRTC for browser or app sessions. Because the pipeline is modular, teams can swap the LLM, voice or language without rebuilding the whole stack.
Pricing is straightforward enough. Voxtral Mini Transcribe 3 costs $0.0035 per minute, Small Transcribe 3 costs $0.004 per minute, Mini Realtime costs $0.008 per minute, and Voxtral Speech 3 is priced at $13 per 1 million characters. The voice agent API is $0.04 per minute with Mistral Small 4, or $0.08 per minute with another reasoning model. OpenRouter also lists the new models with EU-hosted, zero-data-retention endpoints.
My take — AI-written commentary, not fact-checked reporting
This is Mistral doing the sensible thing: sell the whole voice stack, not just another chatbot with a microphone glued on. The real tell is that the company is comfortable bragging about batch transcription and diarization, while the realtime numbers still leave room for the competition to breathe. Europe loves sovereign AI until the latency shows up, and then everyone suddenly becomes very interested in benchmarks.
Read more about this at: MarkTechPost