Mistral AI has introduced Voxtral, a groundbreaking family of open-source speech understanding models. The models are available in two variants: the Small model with 24B parameters, suitable for production deployments, and the Mini model with 3B parameters, designed for edge computing needs. Both versions are released under the Apache 2.0 license, ensuring free accessibility and open use.
Voxtral offers state-of-the-art transcription accuracy, supporting eight languages including English, Spanish, and French. This makes it ideal for businesses needing multilingual support. The models are capable of handling long-form audio, with transcription capabilities extending to 30 minutes and understanding stretching up to 40 minutes. Additionally, Voxtral includes advanced features such as built-in question and answer capabilities, summarization, and function execution triggered by voice commands.
One of the standout aspects of Voxtral is its competitive pricing, starting at just $0.001 per minute. This offers a significant cost advantage compared to proprietary models such as OpenAI’s Whisper, while delivering comparable, if not superior, performance. Mistral positions Voxtral as a robust alternative to closed systems, undercutting their costs by more than half.
For deployment, Voxtral supports private installations and domain-specific fine-tuning, providing flexibility for enterprise applications. It can be accessed through both Hugging Face API and Mistral’s Le Chat, offering easy integration into various platforms. This release marks Voxtral as Mistral AI’s first venture into open-source audio AI models, setting a new standard for innovation and accessibility in the automatic speech recognition domain.

Collections

Mistral AI launched Voxtral, an open-source speech recognition model family with two variants: Small (24B parameters) for production and Mini (3B parameters) for edge computing. Released under Apache 2.0 license, it supports 8 languages, handles up to 40 minutes of audio, and includes Q&A, summarization, and voice command features. Priced at $0.001 per minute, it costs significantly less than proprietary alternatives like OpenAI's Whisper while delivering comparable performance. The models support private deployment, domain-specific fine-tuning, and integration through Hugging Face API and Mistral's Le Chat.

Mistral Introduces Voxtral: An Affordable Open-Source Speech Recognition Model