Gwangju Institute of Science and Technology (GIST, President Lim Ki-cheol) announced on the seventh that the GIST research team won the judges’ award on the ‘International AI Sound Recognition Competition (DCASE Challenge 2024)’ workshop held in Tokyo, Japan from October 23 to 25.
DCASE Challenge 2024 is a world acoustic scene and event detection and classification competition hosted by the Signal Processing Society (AASP) of the Institute of Electrical and Electronics Engineers (IEEE).
The ‘GIST Onion AI’ team, composed of scholars from the Audio Intelligence Lab (AiTeR, Professor Hong-Guk Kim) of the Department of Electrical and Computer Engineering, received support from AunionAI, founded by Professor Hong-Guk Kim. ‘Audio source separation task based on language queries’ After taking first place within the category last July, his performance was recognized by receiving the judges’ award.
The Jury Award is evaluated based on the technical report submitted to the DCASE 2024 Challenge and is reviewed based on the originality and excellence of the research, independent of the challenge rating, and is awarded to just one team per task.
‘Language query-based audio source separation (LASS) technology’ is a technology that separates audio signals in line with the text entered by the user. It provides the idea for developing a generative AI model that connects language and audio, and will be utilized in a wide range of application fields reminiscent of automatic audio editing, multimedia content search, and augmented listening.
The research team announced a high-performance language query-based audio source separation technology through ▲ Large Language Model (LLM)-based prompt technology and data augmentation technology ▲ Fusion technology of pre-learning training models and inference results of existing models ▲ Ensemble technology to enhance AI capabilities did it
Professor Kim Hong-guk said, “The AI model developed through collaboration between the GIST laboratory and Onion AI is critical in that it suggests the opportunity of commercialization quite than staying within the laboratory.” He added, “Particularly, we’re constantly improving the LLM-based audio generation and recognition AI model. “With our efforts, we are going to apply this to varied fields and contribute to the event of technology for a more convenient and safer life,” he said.
Meanwhile, the research was carried out with the support of the MIT International Joint Research Project, the GIST Science and Technology Innovation Project’s ‘Practicalization Research and Development Project’, and the Research and Development Special Zone Promotion Foundation’s ‘Science and Technology Project to Open the Way forward for the Region’ project.
Reporter Park Soo-bin sbin08@aitimes.com
