Mistral Voxtral Small Transcribe 3
Mistral's flagship transcription model, which returns a timestamped transcript and detects the language on its own. In public preview.
Voxtral Small Transcribe 3 is a cloud speech-to-text model from Mistral. It transcribes spoken audio into written text. It can return timestamped segments along with the transcript. It runs through OpenRouter using your own API key, from about $0.004 per minute of audio.
- Released
- Oct 2026
- Pricing
- $0.004 per minute of audio
- Type
- Speech-to-text
- Timestamps
- Timed segments
Samples
Transcribed by Voxtral Small Transcribe 3 via OpenRouter. The same three clips go to every speech-to-text model in the catalog, so the only thing that changes between two models’ transcripts is the model. The clips are synthetic speech made with ElevenLabs v4, and a word that differs from the spoken script is underlined. The numbers clip is left unmarked, since “12” and “twelve” are both right.
Baseline accuracy and punctuation on clean speech
Thanks for joining on such short notice. We moved the launch to Thursday, so could you send me your notes before lunch? I will share the final schedule with everyone tomorrow morning.
Thanks for joining on such short notice. We moved the launch to Thursday, so could you send me your notes before lunch? I will share the final schedule with everyone tomorrow morning.
How dates, amounts, units and a surname are written out
Invoice 4471 is due on March 3rd, 2027. The total comes to $1,249.50 for 12 units, and the parcel weighs 2.5 kilograms. Please call Dr. Elena Markovic at 9:45 a.m. to confirm.
Invoice 4-4-7-1 is due on March 3, 2027. The total comes to $1249.50 for 12 units, and the parcel weighs 2.5 kilograms. Please call Dr. Elena Marković at 9:45 A-M to confirm.
Background noise, a second accent, speaker changes
Excuse me, is this seat taken? No, go ahead. Are you heading to the conference as well? I am. Do you know which stop it is? The one after the river, I think. We can walk over together.
Excuse me, is this seat taken? No, go ahead. Are you heading to the conference as well? I am. Do you know which stop it is? The one after the river, I think. We can walk over together.
Mistral
Mistral is a Paris-based lab that ships fast, efficient open-weight language models with strong reasoning per parameter.
mistral.ai ↗Mistral's largest open-weight model, a multimodal mixture-of-experts built for reasoning, coding and agent work, with vision, tool calling and a 512K-token context window.
Mistral's text-to-speech model with preset voices in eight languages; emotion and pace tags written into the text, such as [emotion=happy] or [pace=slow], direct the delivery. In public preview.
Mistral's compact transcription model, a lower-cost sibling of Voxtral Small Transcribe 3 with timestamps and language detection. In public preview.
Mistral's 24B instruction-tuned model with vision input, robust function calling, and a 128K context.
Mistral's 14B dense edge model (Apache 2.0) tuned for fast on-device reasoning and tool use.
Mistral's efficient 8B dense edge model (Apache 2.0) for on-device chat, reasoning, and tool use.