mistralai / Voxtral-Mini-4B-Realtime-2602
Voxtral-Mini-4B-Realtime-2602
Voxtral Mini 4B Realtime 2602 is a multilingual, realtime speech-transcription model and among the first open-source solutions to achieve accuracy comparable to offline systems with a delay of <500ms. It supports 13 languages and outperforms existing open-source baselines across a range of tasks, making it ideal for applications like voice assistants and live subtitling.
## Overview
Voxtral Mini 4B Realtime is a state-of-the-art streaming speech-to-text model by Mistral AI. It is among the first open-source solutions to achieve transcription accuracy comparable to offline systems with latency under 500ms. The model features a natively streaming architecture with a custom causal audio encoder, enabling real-time transcription as you speak.
## What you can build
- Live transcription and captioning applications
- Real-time voice dictation tools
- Meeting transcription and note-taking
- Voice-to-text input for any application
- Multilingual live subtitling
- Customer support call transcription
## Key features
- Real-time streaming: Transcribes speech token-by-token as audio arrives
- 13 languages: Arabic, Chinese, Dutch, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Spanish
- Configurable latency: Choose between Fast (240ms), Balanced (480ms), or Accurate (960ms) transcription delay
- High accuracy: Matches leading offline transcription models at the recommended 480ms delay
- Efficient architecture: ~3.4B language model + ~970M audio encoder
## Inputs
- Microphone audio: Real-time audio stream from your microphone via WebSocket
- Transcription Delay: Fast (~240ms), Balanced (~480ms, recommended), or Accurate (~960ms)
## Outputs
- Live text transcription: Progressive word-by-word transcription as you speak
- Saved transcript: Complete transcription saved as a text file when the session ends
## Recommended settings
- Use Balanced transcription delay for the best trade-off between speed and accuracy
- Ensure a quiet environment for best transcription quality
- Speak clearly and at a natural pace
- The model auto-detects language automatically
## How it works
1. Click Run to start a realtime session
2. Grant microphone access when prompted
3. Start speaking - transcription appears in real-time
4. Click End Session when finished
5. Session automatically ends after 30 seconds of silence
## Limitations
- Transcription quality depends on audio clarity and background noise
- Very short utterances may not be transcribed
- Silent audio is automatically filtered to prevent hallucination
- Session ends automatically after 30 seconds of no speech activity
## Technical details
- Architecture: Whisper-style audio encoder + Mistral-based language model decoder
- Parameters: ~3.4B (language model) + ~970M (audio encoder)
- Streaming: Native causal attention with sliding window for infinite-length streaming
- **License**: Apache 2.0
API quick start
Run Voxtral-Mini-4B-Realtime-2602 with a single API call.
{
"transcriptionDelay": "balanced"
}