Voxtral Small: speech and audio from Mistral

Voxtral Small (voxtral-small) is Mistral's specialized model for audio and speech. Most language models work only with text, while Voxtral takes sound directly, without a separate speech recognizer.

Why a separate audio model

The classic pipeline "speech recognizer, then text, then language model" loses a lot along the way: intonation, pauses, who is speaking, emotion. Voxtral works with the sound itself, so intonation and pauses make it to the analysis. For calls and meetings this is noticeable.

What to put on it

• Transcription of calls, meetings, voice notes.
• Summaries of recordings: not a verbatim transcript, but "what was agreed and what the action items are".
• Understanding commands in voice interfaces.

How to call it

from openai import OpenAI
client = OpenAI(base_url="https://api.yenisei.ru/v1", api_key="YENISEI_KEY")
resp = client.chat.completions.create(
    model="voxtral-small",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Summarize the call: decisions and action items."},
            {"type": "input_audio", "input_audio": {"data": "<base64>", "format": "mp3"}},
        ],
    }],
)
print(resp.choices[0].message.content)

Russian speech

We tested it separately on Russian calls with background noise and interruptions, the typical reality of a call center rather than a studio recording. Voxtral handles Russian noticeably better than one might expect from a European model: it sometimes alters names and terms, but it conveys the overall meaning and structure of the dialogue correctly.

A practical scenario

The pairing "Voxtral, then a text model" works best: Voxtral turns a call into meaningful text, and then Mistral Small 3.2 (mistral-small-3.2) or another text model writes a brief summary, extracts action items, creates tickets. Each model does its own job, and the result is better than one model for everything.