AI Models Released by META Fundamental AI Research Team
Researchers from the META Fundamental AI Research team have recently unveiled four new AI models targeted towards developers and researchers. These models include JASCO, AudioSEAL, and two versions of Chameleon. One standout among them is the JASCO model, which has been detailed in an article on the Arxiv server.
JASCO: Enhancing Sound Quality and Generating Music
The JASCO model is capable of analyzing various audio recordings to enhance their quality. Users have the ability to adjust the sound of individual instruments like drums, bass guitars, and melodies. Furthermore, JASCO can generate music from scratch based on a text description. META researchers have compared JASCO to similar systems and found that it outperforms competitors in three key metrics.
AudioSEAL: Identifying AI-Generated Speech
The AudioSEAL model is designed to watermark speech produced by AI applications, making it easier to distinguish artificially generated content. It can also label artificial speech segments added to real speech. This model will be available under a commercial license, expanding its potential applications in various commercial projects.
Chameleon: Text-to-Image Conversion
Two versions of the Chameleon model, 7b and 34b, have been created to convert text into visual images. These models will be accessible with limited functionality. The team highlights that both 7b and 34b versions can interpret text and images, enabling reverse processing like generating captions for images.
*META and its products are considered extremist, with their activities prohibited in the Russian Federation.