Скачать или смотреть Continuous Audio Language Models

Researchers have developed a new system for artificial intelligence to create high-quality music and speech called *Continuous Audio Language Models (CALM)**. Older audio models usually turn sound into small, simplified pieces called "tokens," which often results in a loss of sound quality or requires massive amounts of computer power to work properly. **CALM avoids this problem by working with continuous audio information* rather than breaking it into discrete bits, allowing the AI to produce much more detailed and realistic sounds using fewer resources. The system uses a *dual-transformer design* to balance long-term structure with short-term details, and it utilizes a "consistency model" to generate audio in just one or two steps. Because it is so efficient, CALM is significantly *faster than previous models* and can even power tools like **Pocket TTS**, a program small enough to turn text into speech in real-time on a basic laptop.

https://arxiv.org/pdf/2509.06926