← All models

Voxtral Mini Transcribe Realtime 3

Audio Preview Full comparison ↗

mistral/voxtral-mini-transcribe-realtime-3 · by Mistral AI · speech-transcription-model

Pricing — 1 offering(s)

Mistral AI

About this offering: Mistral's realtime speech-to-text model, in Public Preview since 2026-10-20, at $0.008 per minute of audio. It streams over WebSocket with configurable latency (80 ms to 2,400 ms), end-of-turn detection, context biasing and speaker diarization in 35+ languages, per Mistral. Mistral also lists a cached-input rate of $0.0008 per minute, which isn't tracked as a separate price here.

Streaming audio minutes

  • $0.008 / minute Current 2026-10-20 → present

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Voxtral Mini Transcribe Realtime 3 fits into a cost-aware routing setup

See how →

Capability profile

How Voxtral Mini Transcribe Realtime 3 rates across core capability dimensions, with the task-level evidence behind each rating.

transcription accuracy Strong

Mistral reports a 5.7% macro-average WER on 35 FLEURS languages at its 2,400 ms latency setting, rising to 7.4% at 240 ms, against 8.4% for GPT Live Transcribe and 8.7% for Scribe v2 Realtime. Vendor-reported.

language support Moderate

35+ languages, fewer than the 50 of the offline Voxtral Transcribe 3 models, per Mistral; vendor-reported.

speaker diarisation Strong

Speaker diarization in the streaming session, per Mistral's model page; not independently assessed.

real time streaming Strong

WebSocket streaming with configurable latency from 80 ms to 2,400 ms (sub-500 ms in the launch post) and native end-of-turn detection, per Mistral.

custom vocabulary Strong

Context biasing in the streaming session, per Mistral.

Benchmarks

Benchmark Score Config Source
FLEURS 35-language WER, streaming at 2,400 ms latency 5.7 % Mistral's launch post: macro-average WER (CER for JA/KO/ZH) over 35 FLEURS languages; 6.0% at 960 ms, 6.4% at 480 ms, 7.4% at 240 ms. Vendor-reported. source ↗

Operator guidance

Choose Voxtral Mini Transcribe Realtime 3 for streaming transcription in voice agents, where end-of-turn detection and a latency dial matter. It costs about the same as Nova-3 streaming. For batch audio, the offline Voxtral Transcribe 3 models are cheaper and more accurate. Figures are vendor-reported and it's a Public Preview.

Use cases

Limitations

Citations