Official provider results
Voxtral 3 launch evaluation
Mistral reports 4.0% FLEURS word error rate and 18.4% diarization cpWER for Voxtral Small Transcribe 3, the lowest cpWER of the seven systems in its chart.
Source snapshots refreshed
Five measurement lanesVOICE / S2S
Reasoning, task completion, conversation dynamics, experience, and latency stay separate.
Benchmark profile
- Scope
- FLEURS 45-language and 35-language transcription, five-set diarization, Earnings-22 keyphrase recall, and four text-to-speech evaluations
- Primary metric
- Word error rate and cpWER (%, lower is better); keyphrase recall, faithful clips and pronunciation accuracy (%, higher is better); judge score on a 1 to 5 scale
- Owner
- Mistral AI
- Available evidence
- Exact values in the chart data of the October 20, 2026 launch post
What it measures
Voice experience
Is the exchange natural, robust, and responsive?
Latency
How long do responses, tool calls, and full tasks take?
Available results
Batch transcription
Mistral-reported. FLEURS covers 45 languages; "specified" means the language was given, "detected" means it was inferred. cpWER is lower-is-better; keyphrase recall is higher-is-better.
| Model | FLEURS 45 WER · specified | FLEURS 45 WER · detected | Diarization cpWER | Keyphrase recall · with context | Keyphrase recall · without context |
|---|---|---|---|---|---|
| Voxtral Small Transcribe 3 | 4.0% | 4.1% | 18.4% | 94.6% | 81.4% |
| Voxtral Mini Transcribe 3 | 4.5% | 4.6% | 18.6% | Not reported | Not reported |
Realtime transcription on FLEURS 35
Mistral-reported macro-average WER over 35 languages; lower is better. The batch Mini model is an offline baseline.
| Model | Configuration | WER |
|---|---|---|
| Voxtral Mini Transcribe 3 | Offline baseline | 4.0% |
| Voxtral Mini Transcribe Realtime 3 | 2,400 ms delay | 5.7% |
| Voxtral Mini Transcribe Realtime 3 | 960 ms delay | 6.0% |
| Voxtral Mini Transcribe Realtime 3 | 480 ms delay | 6.4% |
| Voxtral Mini Transcribe Realtime 3 | 240 ms delay | 7.4% |
Text to speech
Mistral-reported; higher is better. Voice Code Bench uses 300 English prompts; pronunciation uses 300 English sentences; multilingual uses 377 prompts in 11 languages; expressive English uses 237 prompts.
| Model | Voice Code Bench faithful clips | Pronunciation accuracy | Multilingual judge score | Expressive English judge score | Price per 1M characters |
|---|---|---|---|---|---|
| Voxtral TTS 3 | 79.7% | 86.8% | 3.16 | 3.13 | $13 |
Interpretation limit
All results are Mistral-run. FLEURS macro-averages use CER for Japanese, Korean, Thai, and Chinese. Diarization averages AMI-IHM, AMI-SDM, Switchboard, Switchboard CallHome, and AISHELL-4, with 30-minute AMI and AISHELL versions. Realtime rows trade latency for accuracy; the Mini Transcribe 3 row on FLEURS 35 is an offline baseline. Voice Code Bench is a BeSimple AI test of numbers and names in support calls; the multilingual and expressive scores are judged by gemini-3.8-flash. The launch post and docs disagree on TTS language count (35+ vs 9) and realtime minimum latency (240 ms charted, 80 ms documented). Competitor rows from Mistral charts are not carried here. These results do not enter text-model or voice-agent rankings.