Voxtral

About

Voxtral is Mistral AI's open-source speech understanding platform that goes far beyond simple transcription. Available in two sizes — a 24B production model and a 3B edge model — both released under Apache 2.0. Voxtral outperforms OpenAI Whisper large-v3 across all benchmarks while costing less than half the price of comparable APIs. Handles audio up to 30 minutes for transcription and 40 minutes for understanding with a 32K token context window. Features built-in Q&A and summarization directly on audio content, automatic language detection across dozens of languages, speaker diarization, and function-calling straight from voice — enabling spoken commands to trigger backend workflows without intermediate parsing. The API routes to Voxtral Mini Transcribe, optimized for cost and latency. Ideal for meeting intelligence, call center analytics, podcast processing, and voice-driven applications.

Features

• •State-of-the-art transcription accuracy beating Whisper large-v3
• •Two model sizes: 24B (production) and 3B (edge/local)
• •32K token context — up to 30min transcription, 40min understanding
• •Built-in audio Q&A and summarization without chaining models
• •Automatic language detection and multilingual support
• •Speaker diarization for multi-speaker audio
• •Function-calling from voice for workflow automation
• •Real-time streaming with sub-200ms latency
• •Apache 2.0 open-source license
• •Less than half the cost of comparable closed APIs

Use Cases

• •Transcribing meetings with speaker identification and action item extraction
• •Building voice-driven assistants that trigger backend workflows
• •Processing podcast episodes with automated summaries and Q&A
• •Call center analytics with multilingual transcription at scale
• •Edge deployment of speech understanding on devices with the 3B model