In this work, we present an approach for acoustic scene classification (ASC) based on machine learning-based sound-to-light conversion implemented on Blinkies. Blinkies are compact devices we have developed that include microphones, LEDs, a small computer, and a battery, and are capable of converting sound into light. Our concept is to distribute Blinkies across a wide area and capture their emitted light using a video camera, making it easy to collect acoustic information from a broad spatial region. This acoustic information can then be utilized for ASC. However, due to the frame rate limitations of standard video cameras, how sound is converted into light becomes critical. To address this, we apply machine learning to the sound-to-light conversion process in order to improve ASC performance. We demonstrate the feasibility of this approach through real-world experiments using a small ASC dataset we recorded. Work supported by JST SICORP Grant No. JPMJSC2306.
Kotsugi et al. (Wed,) studied this question.