Mistral releases Voxtral 3 speech models for voice agents

• Voxtral 3 ships transcription, speech generation and a LiveKit-based voice-agent SDK, introduced in MistralAI's announcement.
• Transcription lands nearer the top of independent leaderboards: ArtificialAnlys placed Small Transcribe third at 2.1% word error, while TTS ranked 12th in both voice arenas.
• Pro Voice adapts a speaker's voice across supported languages using roughly 30 minutes of audio, according to the voice-adaptation announcement.
• API pricing starts at $0.0035 per audio minute for batch transcription and $13 per million TTS characters, as WesRoth's launch summary reported.

Word timestamps are the migration landmine: Mistral's migration guide says requesting them now returns a 400. The voice-agent defaults also include a 500 ms minimum-silence setting and allow callers to interrupt mid-sentence.

The SDK orchestrates three separate models as one streaming pipeline:

1. Listen: Voxtral Mini Transcribe Realtime 3 transcribes speech and detects turn endings.
2. Reason and act: A general-purpose LLM generates answers and calls tools.
3. Speak: Voxtral Speech 3 streams the spoken response.

Models and voices can be swapped without rewriting the agent. Phone calls use SIP, browser and app connections use WebRTC, and tools support function calling, web search and retrieval over business systems, according to Mistral's launch post.

Small Transcribe supports speaker diarization and vocabulary biasing, as described in the transcription announcement. Both batch models cover 50+ languages; Mini is a 4B-parameter model.

Artificial Analysis's launch-day results were:

• Small Transcribe 3: 2.1% AA-WER, ranked third.
• Mini Transcribe 3: 2.2% AA-WER, ranked fifth.

AA-WER v2.5 weights AgentTalk at 50%, VoxPopuli at 25% and Earnings22 at 25%, according to the evaluator's methodology.

The realtime model combines acoustic and semantic signals to predict when a speaker has finished, according to Mistral's technical description. Mistral claims conversational turns under one second for the complete agent pipeline.

For transcription alone, ArtificialAnlys measured two settings:

• 480 ms: 3.6% final-transcript AA-WER, first partial transcript in 0.27 seconds, final transcript 0.75 seconds after speech ends.
• 960 ms: 3.3% final-transcript AA-WER, first partial in 0.23 seconds, final transcript 1.1 seconds after speech ends.

The previous Voxtral Mini Transcribe Realtime scored 5.2%, making those reductions 1.6 and 1.9 percentage points, respectively. Transcript-finalization latency and full-agent response latency measure different events.

Voxtral Mini TTS 3 supports 35+ languages and 50+ dialects, with two delivery-control paths:

• Implicit: Emotion, pace and emphasis follow the conversation's meaning.
• Explicit: Inline tags such as [emotion=sad][pace=slow] set delivery line by line.

Pro Voice uses roughly 30 minutes of recorded audio to reproduce a speaker's character, emotional range and speaking style across supported languages.

Artificial Analysis measured provider-voice Elo at 1,218, ranking 12th of 98 models. Controlled-voice Elo was 1,068, ranking 12th of 43, with the same eight cloned voices used across systems.

Pronunciation robustness reached 84.4%, ranking 13th of 32 models. Human reviewers scored four categories:

• Standalone terms: 93.9%.
• Contextually appropriate pronunciation: 91.9%.
• Expanding shorthand: 80.2%.
• Preserving exact sequences: 75.5%.

Mistral's launch price list separates component APIs from the voice-agent API:

| Service | Price |
| --- | --- |
| Mini Transcribe 3, batch | $0.0035/minute |
| Small Transcribe 3 | $0.004/minute |
| Mini Transcribe Realtime 3 | $0.008/minute |
| Speech 3, batch or realtime | $13/million characters |
| Voice agent with Mistral Small 4 reasoning | $0.04/minute |
| Voice agent with another reasoning model | $0.08/minute |

The availability section lists European, US and global API endpoints. For on-premises or private-cloud deployment, Mistral directs customers to contact its team.

OpenRouter added the two batch transcribers and TTS model, with EU-hosted, zero-data-retention endpoints using /audio/transcriptions and /audio/speech.

The migration guide documents several integration changes:

• Word timestamps disappear: timestamp_granularities accepts segment only; requesting word returns 400.
• Realtime settings lock after audio starts: A later session.update produces an error event.
• Diarized segments can overlap: Transcript displays can no longer assume non-overlapping segments.
• The TTS alias will move: voxtral-mini-tts-latest is scheduled to point to Voxtral 3 within a month of general availability.
• Free-form style instructions are unavailable at launch: The guide lists them among new TTS features but explicitly marks them as not yet available.