Mistral Voxtral Mini Transcribe Realtime

Overview

Best for: EU-vendor realtime STT with an escape hatch to run the identical model on your own GPUs.

At a glance

Latency configurable down to sub-200 ms (vendor). $0.006/min from the launch announcement, not a live price table. Open weights Apache 2.0 (Voxtral-Mini-4B-Realtime). Realtime cannot be combined with diarization.

Features

• streaming text deltas
• configurable latency
• short-lived rt_ tokens for browsers
• open weights for on-prem

Pricing

Free tier: Not stated.

Source: mistral.ai

Setup

1. pip install "mistralai[realtime]>=2" (V2 SDK).
2. Stream PCM to client.audio.realtime.transcribe_stream with model voxtral-mini-transcribe-realtime-2602.
3. For browsers: server calls POST https://api.mistral.ai/v1/client/sessions {purpose:'realtime', model} to mint an rt_ token; browser opens the WebSocket with subprotocols ['realtime', token].
4. Self-host: download the HF weights and serve with an engine that supports the realtime model (check model card).

Endpoint

wss://api.mistral.ai/v1/audio/transcriptions/realtime?model=voxtral-mini-transcribe-realtime-2602

Authentication

Authorization: Bearer <key> server side; rt_ short-lived token via Sec-WebSocket-Protocol in browsers.

import asyncio, os  # pip install "mistralai[realtime]>=2"
from mistralai.client import Mistral
from mistralai.client.models import (AudioFormat, TranscriptionStreamDone,
                                     TranscriptionStreamTextDelta)

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
fmt = AudioFormat(encoding="pcm_s16le", sample_rate=16000)

async def audio_chunks():
    with open("audio_16k_mono.raw", "rb") as f:
        while chunk := f.read(15360):  # 480 ms
            yield chunk
            await asyncio.sleep(0.48)

async def main():
    async for ev in client.audio.realtime.transcribe_stream(
        audio_stream=audio_chunks(),
        model="voxtral-mini-transcribe-realtime-2602",
        audio_format=fmt,
    ):
        if isinstance(ev, TranscriptionStreamTextDelta):
            print(ev.text, end="", flush=True)
        elif isinstance(ev, TranscriptionStreamDone):
            print("\n[done]")

asyncio.run(main())

Written from the current docs. Check the vendor's SDK version before you ship.

Warnings

Plus 12 warnings that apply to all speech-to-text, live APIs. See category warnings.

Limits

• rt_ client tokens expire after about 900 seconds; mint them just before connecting.
• Python realtime needs SDK v2 (mistralai>=2, extra [realtime]); not available in v1 SDK.

Models and products

Docs and sources

Docs

• Realtime transcription docs
• Client authentication
• Announcement
• Open weights

Sources used

• docs.mistral.ai/studio/audio/speech_to_text/realtime_transcription.md
• docs.mistral.ai/studio/audio/speech_to_text/realtime_transcription/client_auth....
• docs.mistral.ai/capabilities/audio_transcription
• mistral.ai/news/voxtral-transcribe-2
• the-decoder.com/voxtral-transcribe-2-offers-speech-recognition-at-0-003-per-min...
• huggingface.co/api/models/mistralai/Voxtral-Mini-4B-Realtime-2602