tech
Voxtral
Open-source ASR systems with high word error rates and limited semantic understanding
TL;DR
- Voxtral models (24B and 3B variants) offer state-of-the-art speech understanding.
- They are released under the Apache 2.0 license and available via API.
- Voxtral provides high accuracy and semantic understanding at a lower cost than comparable proprietary APIs.
- Key capabilities include long-form context handling, built-in Q&A and summarization, multilingual support, function-calling, and strong text understanding.
- Voxtral outperforms existing models like Whisper, GPT-4o mini, and Gemini 2.5 Flash on various benchmarks.
- The models can be downloaded locally, accessed via API, or tested on Le Chat.
- Enterprise features include private deployment, domain-specific fine-tuning, advanced context options, and dedicated integration support.
- Future updates will include speaker segmentation, audio markups, word-level timestamps, and non-speech audio recognition.