Introduction to Audio at Mistral AI

Course updated November 3, 2026 for Voxtral 3, based on the Mistral guide Migrate from Voxtral 2: 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), real-time diarization supported, 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.

The Voxtral Family: Mistral AI’s Audio Suite

Mistral AI has moved beyond text. With the Voxtral family, the French company offers a complete suite of models dedicated to speech: transcription, audio understanding, and text-to-speech. This course teaches you to leverage these capabilities to turn sound streams into truly usable text, complete with the accompanying Python code.

Automatic speech recognition (ASR), or Speech-to-Text (STT), powers everyday tools you likely use. Subtitles on training videos, executive meeting minutes, customer service voice assistants, and digital accessibility on public websites all rely on the same core technology. Mistral AI brings a sovereign, high-performance, and privacy-respecting approach to this space—critical when your audio contains patient names, customer data, or internal deliberations.

Two Ways to Transcribe

Voxtral supports two modes that address entirely different constraints. The first is offline transcription, also known as batch processing: you send a complete audio file and receive the transcription. This mode fits recorded podcasts, field interviews uploaded later in the day, or archived lectures. Voxtral Mini Transcribe 3 and, for higher accuracy, Voxtral Small Transcribe 3 handle this, offering diarization and segment-level timestamping.

The second is real-time transcription, or streaming: audio is transcribed as it arrives, with a configurable latency between 80 ms and 1.2 seconds (or 2.4 seconds). Nothing is pre-recorded; everything happens while the person speaks. This mode powers live webinar subtitles, voice assistants, or dictation tools. The Voxtral Mini Transcribe Realtime 3 model is built for this. Keep this distinction in mind now: it drives most of your technical choices throughout the rest of the course.

Supported Languages

Voxtral 3 file transcription models cover 50 languages, while the real-time model supports 45 languages and dialects, including French. Without a language parameter, the model auto-detects the recording’s language; if you know it, specify it using its two-letter code (e.g., fr).

What Voxtral Does Beyond Raw Text

Transcription alone isn’t always enough. Diarization automatically identifies speakers: in a five-person meeting, the model distinguishes who spoke and when, turning a wall of text into a readable summary. Context biasing lets you inject specialized vocabulary—proper names, technical terms, acronyms—so “Kubernetes” or “GDPR” don’t get mangled in an engineering meeting.

Segment-level timestamping provides the start and end time for each transcription segment, enabling synchronized subtitle generation; Voxtral 3 no longer supports word-level timestamping. Noise robustness maintains accuracy in challenging acoustic environments: background noise, meeting room reverberation, or laptop microphones. Finally, you can submit recordings as files, via URL, or from previously uploaded files, and receive responses as a stream, chunk by chunk.

Compliance and Sovereignty

The previous generation of the real-time model, Voxtral Realtime (voxtral-mini-transcribe-realtime-2602, 4 billion parameters), is released as open weights under the Apache 2.0 license: you can self-host it without relying on an external API, even after its retirement from the API on April 30, 2027.

This is a major advantage for European companies facing strict regulatory constraints on audio data processing: meeting recordings, customer calls, or medical consultations never leave your infrastructure.

What You Will Learn

The course follows a straightforward progression. You’ll start by understanding the three Voxtral models and choosing the right one for your use case. Next, you’ll tackle offline transcription, covering the API, its parameters, diarization, and context biasing. Then you’ll move to real-time transcription: streaming, microphone capture, and latency control. You’ll finish with production deployment, including optimization, error handling, and cost management. Every lesson includes complete, functional Python code that you can run as-is.

Key Takeaways

• Voxtral is Mistral AI’s audio model family, covering STT and TTS
• Two transcription modes: offline (files) and real-time (streaming, configurable latency starting at 80 ms)
• 50 languages for files, 45 for real-time, including French
• Advanced features: diarization, context biasing, segment-level timestamping, noise robustness
• Sovereign deployment possible with the second-generation real-time model, available as open weights under Apache 2.0