AiOla open-sources ultra-fast 'multi-head' speech recognition model

aiOla has unveiled Whisper-Medusa, an innovative open-source AI model that surpasses OpenAI’s Whisper by achieving 50% faster performance without sacrificing accuracy. The advancement comes from its ability to predict ten tokens at once, significantly increasing speech prediction speed and enhancing runtime, particularly for long-form audio. Whisper-Medusa is currently offered as a 10-head model, with future plans to introduce a 20-head version.

• Combines OpenAI's Whisper with aiOla's technology, over 50% faster.
• Whisper-Medusa capable of predicting ten tokens at a time.
• Available on Hugging Face and GitHub for public access.
• Employs multi-head attention architecture and weak supervision.
• Understands over 100 languages and various business jargons.