Mistral launches Voxtral 3 voice models with Emirati Arabic preset

The latest:

Mistral released Voxtral 3 on October 20, 2026, a four-model speech family built around transcription, realtime streaming and text-to-speech, with a voice-agent SDK the company says completes a full turn in under one second. Pricing starts at $0.0035 per minute for batch transcription and $0.04 per minute for the agent stack. All four models entered public preview, with self-deployment offered on customer infrastructure.

Details:

• The lineup: The family includes Voxtral Small Transcribe 3 for highest-accuracy work, Voxtral Mini Transcribe 3 at 4 billion parameters for batch jobs, Voxtral Mini Transcribe Realtime 3 for live streaming, and Voxtral Mini TTS 3, listed on the price sheet as Voxtral Speech 3. Mistral did not disclose the Small model’s size.
• Accuracy claims: On Mistral’s own FLEURS testing across 45 languages, Voxtral Small Transcribe 3 posted roughly 4.1% word error with specified language and 4.0% inferred, matching Microsoft AI Transcribe 2. The Mini model scored about 4.6%, ahead of ElevenLabs Scribe v2 at 4.9%, Gemini 3.5 Transcribe at 6.1% and Whisper large-v3 at 16.0%.
• Arabic coverage: FLEURS includes Arabic alongside Chinese, English, French, German, Hindi, Japanese, Korean, Spanish and Vietnamese. The speech side covers more than 35 languages and over 50 accents, and Mistral lists Arabic in its voice library with two presets: Modern Standard Arabic and Emirati.
• Speaker separation: All transcription models support speaker diarization and context biasing, where clients supply terms or free text to sharpen recognition. On cpWER across five public datasets, Voxtral Small scored 18.4% and Mini 18.6%, ahead of Pyannote at 19.0%, Microsoft at 20.3% and Gemini 3.5 Transcribe at 38.4%.
• Context payoff: On the Earnings 22 benchmark, Mistral reported keyphrase recall of 94.6% for Voxtral Small Transcribe 3 when context was supplied, against 81.4% without it — a 13-point gap that matters for dialect-heavy or jargon-heavy calls where names and product terms drive the transcript.
• The latency tradeoff: Realtime transcription runs at selectable latencies of 240, 480, 960 and 2400 milliseconds. Mistral reported about 5.7% word error at 2400 ms, 6.4% at 480 and 7.4% at 240 across 35 FLEURS languages. Gemini 3.5 Realtime posted 4.2%, Microsoft Streaming 5.8%, GPT Live Transcribe 8.4%.
• Voice controls: Voxtral Speech 3 offers 34 emotion tags plus pacing and emphasis control. A Pro Voice option adapts a speaker’s style from roughly 30 minutes of audio, the same threshold used for voice cloning. Mistral says the model undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash on cost, without publishing comparison figures.
• The agent stack: The SDK is built on LiveKit and chains realtime transcription with end-of-turn detection, reasoning and tool calls through a general model such as Mistral Small 4, then speech output. Each stage is a separate model, so teams can swap model, voice or language without rebuilding the agent.
• Telephony hooks: The SDK supports phone calls over SIP and browsers and apps over WebRTC, and connects to IVR systems and contact-center software. Agents stop speaking when a caller interrupts. Listed use cases include caller identity verification, bill checks and appointment booking, with a demo translating a French medical referral.
• Price ladder: Batch transcription costs $0.0035 per minute on Mini and $0.004 on Small; realtime runs $0.008 per minute; speech output is priced at $13 per million characters. The agent interface costs $0.04 per minute with Mistral Small 4 and $0.08 with a larger reasoning model. Self-deployment goes through sales.

Between the lines:

The 240-millisecond setting carries a measurable accuracy cost, rising from 5.7% to 7.4% word error, so the sub-second agent turn Mistral advertises is a deliberate tradeoff rather than a free gain. The modular pipeline and the self-deployment option point at regulated buyers — banks, hospitals, telecom operators — who need Arabic voice handling inside their own infrastructure rather than on a vendor endpoint.

What’s next

The four models sit in public preview on European, American and global endpoints, with testing playgrounds in Studio. No licensing terms or open weights were announced, and general availability dates were not set.