Tasks and inputs
Voxtral Mini Transcribe Realtime 3 is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents: under 500 ms, tunable delay, built-in turn detection.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Indicative: $0.008 per audio minute. Voice agent API with Mistral Small 4 about $0.04/min, larger model $0.08/min. Confirm on your own account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
How long can a caller wait before a voice agent feels broken? Somewhere under a second, in most people's experience, and transcription is only the first hop. Voxtral Mini Transcribe Realtime 3 is Mistral's streaming recogniser for exactly that budget: it targets latency below 500 ms and lets you choose where on the speed-versus-accuracy curve each session sits.
Released on 20 October 2026, it replaces voxtral-mini-transcribe-realtime-2602, which was deprecated on 31 October.
Streaming delay is set per session. Allowed values run from 80 ms to 1,200 ms, plus a 2,400 ms option. Shorter delay means the model commits to words with less lookahead, so expect more corrections on fast or mumbled speech. Longer delay gives it more acoustic context before it commits.
A sensible pattern: a short delay for barge-in-heavy phone agents, a longer one for live captioning where a little lag is acceptable and accuracy is what viewers notice.
The model predicts end-of-turn natively, using both acoustic cues and the meaning of what was said. That matters. A pause after "my account number is" is not the end of a turn, and a voice-activity detector that only listens for silence will cut the caller off. Having this inside the recogniser removes one component from your stack, though you should still test it on your callers' speech patterns.
| Item | Detail |
|---|---|
| API ID | voxtral-mini-transcribe-realtime-3-0 |
| Mode | Streaming over WebSocket |
| Latency target | Under 500 ms |
| Delay setting | 80–1,200 ms, or 2,400 ms |
| Languages | 35+ (the batch models cover 50+) |
| Indicative price | About $0.008 per audio minute |
| Weights | Closed |
The Python SDK's realtime client installs with the mistralai realtime extra (pip install "mistralai[realtime]").
Mistral also sells a voice agent API built on LiveKit that wires speech-to-text, a reasoning model and text-to-speech into one streaming pipeline, with SIP for phone lines and WebRTC for browsers. Indicative pricing is $0.04 per minute when the reasoning step uses Mistral Small 4, and $0.08 with a larger model. If you would rather own orchestration, call this recogniser directly and pick your own LLM.
Note the language gap. With 35+ languages against the batch tier's 50+, check that every locale you serve is supported before you route live traffic.
Record real sessions, with consent, and replay them through the WebSocket at real-time speed. Measure three things: time from end of speech to final transcript, how often a turn is cut early, and how often it is held too long. Try at least two delay settings. An alternative worth benchmarking alongside is GPT Realtime Whisper; more options sit under speech to text.
At roughly double the batch price per minute, this model is for audio that has to be understood now. Anything that can wait should go through batch. Confirm prices on your own account.
Compare the model on the work you need to complete.
Voxtral Mini Transcribe Realtime 3 is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
salesforce
SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.
ElevenLabs
Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.
NVIDIA
Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.
IBM
Granite Speech 5.0 2B is IBM's open speech-language model for multilingual ASR and translation with keyword biasing, runnable from Ollama to MLX.
Mistral AI
Voxtral Mini Transcribe 3 is Mistral's 4B batch transcription model with speaker labels, segment timestamps and custom vocabulary, replacing Mini 2.
Mistral AI
Voxtral Small Transcribe 3 is Mistral's top-accuracy batch speech-to-text model for recorded audio, with context biasing and 50+ languages.