“AI Sound Analysis: Endless Applications”
관리자
2024-11-11
Suji Lee (right), CEO of Deeply, and Myunghoon Ryu, CIO.
Deeply, a startup specializing in AI sound detection and analysis solutions is making strides in the field. In an interview with The Korea Economic Daily, CEO Suji Lee emphasized, “The potential of sound AI is immense—from detecting abuse situations in call centers to monitoring livestock health on farms.”
Deeply’s proprietary sound detection system, ListenAI, utilizes a dataset of 42 different sound types, totaling 50,000 hours of audio. This enables it to identify sudden incidents across various environments. For example, it can detect machine malfunctions in manufacturing facilities or alert workers to fall accidents at construction sites through sounds like crashing or shouting.
While existing AI services focused on voice analysis for translation, specialized AI for sound classification and analysis has been rare. CIO Myunghoon Ryu noted, “We have developed an integrated solution combining AI software with specialized hardware like microphones and edge servers.”
Deeply collaborates with major partners, including Kangwon Land, Lotte Construction, and KORAIL, to implement its solutions effectively. More recently, ListenAI has been implemented at facilities like the Government Sejong Complex Sports Center and restrooms at Incheon National University Station. The system detects sounds in real-time and issues alerts upon identifying abnormal situations. CEO Lee stated, “Even where CCTV is installed, incidents can be missed without continuous monitoring. ListenAI enhances the accuracy of surveillance when used alongside CCTV.” KORAIL also employs the solution to monitor the sounds of traction motors for early fault detection.
Lee founded Deeply after identifying the technical challenges in sound AI compared to visual AI. He said, “For sound AI to function effectively, it needs to account for factors like echo, noise, and human speech characteristics. This makes it a field that is not easily replicable by other AI companies.”
To develop their sound AI models, Deeply gathered diverse sound data, including screams, fighting noises, and object impacts. Ryu shared, “We acquired external data, hired actors for specific recordings, and curated a dataset tailored to sound event lists for training models.”
Deeply has also embarked on a project with Korea University Hospital to track the daily routines of depression patients, monitoring activities such as regular exercise and meals through sound analysis.
“The possibilities for sound AI are endless,” said Lee. “Deeply aims to be at the forefront of this pioneering field.”
— Euni Ko, The Korea Economic Daily