Mistral's Voxtral 3 speech models benchmarked: top-3 STT, mid-pack TTS

Mistral AI has released Voxtral 3, one voice-generation model and three transcription models, and Artificial Analysis has run all four through its speech leaderboards (Artificial Analysis on X). The split result is that Mistral's transcription is close to the best available, while its synthetic voice is competitive but not a leader.

On transcription, Artificial Analysis ranks models by its AA-WER Index, a word error rate averaged across three test sets, one of them a held-out collection of voice-agent conversations, where word error rate counts the words a model substitutes, adds or drops relative to a reference transcript (methodology). Lower is better. Voxtral Small Transcribe 3 scores 2.1%, third on the board and just ahead of ElevenLabs Scribe v2 at 2.2% and Grok Voice Transcribe 2.0 at 2.3%, and the smaller Voxtral Mini Transcribe 3 is fifth at 2.2%. The realtime model, which transcribes audio as it streams in, scores 3.3% on its final transcript at the 960 ms setting, sixth of 46, and 3.6% at 480 ms, down from 5.2% for the previous Voxtral Mini Transcribe Realtime.

Voxtral 3 TTS sits further back on the listener-preference boards, which rank models by Elo from votes in Artificial Analysis's speech arena. With each provider's own voices it scores 1,218, twelfth of 98 and within a few points of Breeze TTS 2 above and two models below. On the Controlled Voice board, where every model speaks the same 8 cloned voices, it is twelfth of 43 at 1,068, level with Ant Group's Ming-3.0-tts and just behind Eleven v3. On Pronunciation Robustness, a test of saying difficult text correctly, it scores 84.4%, thirteenth of 32. Artificial Analysis lists its price at $13 per million characters, below Gemini 3.8 Flash TTS at $16.5 and Eleven v4 Turbo at $40, though above Simba 3.2 at $6.6.