"Deepening Sound Analysis Capabilities with AI": Deeply Publishes Research at International Conference
관리자
2024-07-24
Deeply, a company specializing in sound analysis artificial intelligence engines, announced on the 23rd that it has published a research paper at ICASSP 2024, an international signal processing conference.
ICASSP, organized annually by the Institute of Electrical and Electronics Engineers (IEEE), is the world's largest international conference in the fields of speech, image, communication, and signal processing. This year marks the 49th conference and, for the first time, it was held in South Korea. The event attracted global companies like Apple and Google, along with over 4,000 researchers.
Deeply’s paper demonstrates advancements in sound analysis AI, particularly addressing the issue of decreased accuracy due to frequency characteristic differences across various recording devices. The solution uses a proprietary acoustic transformation technology to generate data for AI training, which has led to improvements in generalization performance (the AI's ability to analyze new data) from 5.2% to 11.5%.
The paper highlights an acoustic transformation model built on the 'CycleGAN', a deep learning algorithm for modifying the attributes of media such as images, sound, and text. The model was further refined with Deeply’s unique data augmentation techniques.
To address variability in recording equipment, Deeply trained the model on the acoustic properties of various devices, applying different frequency patterns to their existing sound data. This approach allowed them to simulate the effect of recording the same sound on multiple devices, enriching the data with diverse frequency patterns.
Deeply's sound analysis AI engine, after refinement using the generated data, maintains high accuracy even in new environments. The engine has been trained on over 50,000 hours of sound data to reach high levels of precision.
Already recognized for its ability to detect subtle sounds, the technology is gaining attention in security sectors where intruder detection is critical. It is also applied in the manufacturing industry for detecting equipment malfunctions and product defects.
Suji Lee, CEO of Deeply, stated, “Our AI sound analysis engine’s accuracy ranks among the highest globally. We aim to continue expanding our capabilities, establishing Deeply as a leader in sound analysis solutions that serve diverse needs across society.”
[Unicorn Factory, MoneyToday Startup Media Platform]