機械学習を用いた音響物理特性に基づく特徴量抽出

抄録

In recent years, there has been a growing need for sound branding that takes advantage of the characteristics of sound. However, sound design is often based on subjective perceptions, and there are few examples of sound design based on the physical characteristics of sound. Additionally, it is difficult to find commonalities between different types of products only by extracting sound features based on frequency characteristics. In this study, a machine learning model for acoustic multi-parameters, based on CNN, was constructed to classify sound data by inputting the results of time frequency analysis, the correlation between carrier frequency and amplitude modulation frequency (Fast-SC), and a psychoacoustic evaluation scale. Then, feature extraction of the parameter with higher contribution to classification is performed for each data using SHAP. The results showed that loudness was higher when the sound pressure was higher, roughness, fluctuation strength and Fast-SC were higher when there was amplitude modulation, and prominence ratio was higher when there were multiple pure tones. This result confirms the effectiveness of feature extraction using a machine learning model for acoustic multi-parameters.