Voxtral Models for Transcription

Course updated November 3, 2026 for Voxtral 3, based on Mistral’s Migrate from Voxtral 2 guide: Voxtral 2 models (voxtral-mini-2602, voxtral-mini-transcribe-realtime-2602, voxtral-mini-tts-2603), deprecated on October 31, 2026 and retired on April 30, 2027, are replaced in examples by voxtral-mini-transcribe-3, voxtral-mini-transcribe-realtime-3, and voxtral-mini-tts-3-0. Key changes: timestamping by segment only (word returns a 400 error), diarization possible in real time, real-time session settings locked after the first audio chunk, and free-form context_bias inputs (phrases with spaces). Sources: Migrate from Voxtral 2 · Speech to Text.

Four models, three use cases

Mistral AI offers several Voxtral models for speech processing, each optimized for a specific task: two for files, one for real time, and one for asking questions about audio. Choosing the right model isn’t a configuration detail; it determines what you can do next—timestamping, turn detection, streaming responses. Make this decision first, before writing a single line of code.

Voxtral Mini Transcribe 3 — file transcription

This is the default model for offline transcription: you send it a complete recording, and it returns the transcribed text. Its API identifier is voxtral-mini-transcribe-3 (an alias for voxtral-mini-transcribe-3-0).

It supports all offline features: diarization identifies speakers with segment-level timestamping (word-level timestamping no longer exists in Voxtral 3), context biasing injects domain-specific vocabulary, and responses can stream. It covers 50 languages and costs $0.0035 per minute of audio as of November 3, 2026.

In practice, choose this model whenever the audio already exists as a file: meeting recordings, podcasts, interviews, lectures, or medical and legal transcriptions.

Minimal call example

from mistralai.client import Mistral

client = Mistral(api_key="YOUR_API_KEY")

with open("meeting.mp3", "rb") as f:
    response = client.audio.transcriptions.complete(
        model="voxtral-mini-transcribe-3",
        file={"content": f, "file_name": "meeting.mp3"}
    )

print(response.text)

Voxtral Small Transcribe 3 — when precision matters

Larger and more accurate, voxtral-small-transcribe-3 works the same way. Mistral recommends starting with the Mini model and switching to Small only if errors persist despite using custom vocabulary. It costs $0.004 per minute.

Voxtral Mini Transcribe Realtime 3 — real-time transcription

This model is designed for streaming transcription. Its API identifier is voxtral-mini-transcribe-realtime-3.

Its latency is adjustable from 80 ms to 1.2 seconds, or 2.4 seconds (target_streaming_delay_ms): the architecture transcribes as audio arrives rather than waiting for the end of a sentence. It covers 45 languages and dialects, detects turn endings, and can now distinguish speakers—but not both in the same session. Its segments include speaker labels but no start or end times. It costs $0.008 per minute. Use it for live captions, voice assistants, dictation, or transcribing a conference as it happens.

Minimal call example

from mistralai.client import Mistral
from mistralai.client.models import AudioFormat

client = Mistral(api_key="YOUR_API_KEY")
audio_format = AudioFormat(encoding="pcm_s16le", sample_rate=16000)

async for event in client.audio.realtime.transcribe_stream(
    audio_stream=my_audio_stream,
    model="voxtral-mini-transcribe-realtime-3",
    audio_format=audio_format,
):
    print(event.text, end="", flush=True)

The previous generation, voxtral-mini-transcribe-realtime-2602 (4 billion parameters), remains available as open weights under the Apache 2.0 license for self-hosting on your own machines.

Voxtral Small — multimodal audio chat

The third model changes categories. Voxtral Small, identifier voxtral-small-latest—still a generation 2 model at the launch of Voxtral 3, according to the documentation—is not a pure transcriber but a conversational model capable of understanding audio. It accepts an audio file as input in a chat conversation and answers your questions about its content, combining audio understanding with text generation.

This difference shows up in the code: it uses the /chat/completions endpoint, not /audio/transcriptions. You won’t ask it “transcribe this file” but rather “summarize this conversation,” “what budget was mentioned,” or “did the customer express a complaint.”

Call example

import base64
from mistralai.client import Mistral

client = Mistral(api_key="YOUR_API_KEY")

with open("audio.mp3", "rb") as f:
    audio_base64 = base64.b64encode(f.read()).decode("utf-8")

response = client.chat.complete(
    model="voxtral-small-latest",
    messages=[{
        "role": "user",
        "content": [
            {"type": "input_audio", "input_audio": audio_base64},
            {"type": "text", "text": "Summarize this conversation in 3 key points."}
        ]
    }]
)

print(response.choices[0].message.content)

Choose in one minute

When facing a new project, ask yourself about the dominant need rather than the available features. The table below summarizes the most common scenarios.

A common mistake is choosing the real-time model because it seems more modern, only to discover late in the project that the expected report requires dating each intervention: in real time, segments include the speaker but not the time. If your deliverable must state “who said what, and when,” offline mode remains the right choice.

Key takeaways

• Voxtral Mini Transcribe 3: file transcription with diarization, context biasing, and segment-level timestamping
• Voxtral Small Transcribe 3: same capabilities, higher precision when custom vocabulary isn’t enough
• Voxtral Mini Transcribe Realtime 3: streaming with adjustable latency from 80 ms, turn detection, diarization without timestamping
• Voxtral Small: audio chat to understand and analyze recording content
• 50 languages for files, 45 for real time, including French