Wavelet-based feature extraction for time-varying sound

We present a wavelet-based feature extraction approach for evaluating sound quality of time-varying sounds. Triangular wavelets that are two-dimensional nonseparable wavelets are used for extracting features in both local and global region of the sounds. Various kinds of button sounds represented in the time-frequency plane are transformed into a multi-scale pyramid consisting of three isotropic wavelet components along the horizontal, vertical and diagonal direction at each scale. We show that the local characteristics of the button sounds as well as global ones can be detected from the certain wavelet components. The relation between the features and the corresponding auditory impressions extracted by the semantic differential method is discussed.