Thanks to the integration of CLaMP and SymFormer neural networks into GigaChat
The SberBank team will soon significantly expand the capabilities of the proprietary generative neural network GigaChat. GigaChat users will be able to generate music tracks based on arbitrary text queries. Denis Filippov, vice president for digital surfaces “Salyut” of Sberbank, spoke about this at the international conference on artificial intelligence AI Journey 2023.
To generate music, a person will only need to formulate a task, for example: “Compose fun country music” or “Write a track for the lounge area of a business center.” In response, GigaChat will provide an audio file with a unique musical composition and a musical score in MIDI format, which any DAW (Digital Audio Workstation) can understand.
The user will be able to listen and download the resulting track, and use the MIDI file in their own creative projects.
Generating music in GigaChat is possible thanks to the integration of the CLaMP and SymFormer neural networks. To train SymFormer, we used the ML Space platform based on the Christofari supercomputer and a dataset of more than 200 thousand compositions of different styles: from classical to modern electronic music and rock.
Denis Filippov said:
The new features of GigaChat will be useful not only for music enthusiasts and representatives of creative professions. We see representatives of medium and small businesses as one of the segments of the service’s target audience. Thanks to GigaChat, they will be able to quickly, efficiently and, importantly, absolutely legally solve business problems: create background music for cafes, beauty salons and waiting rooms, generate tracks for advertising videos and social networks.