Similarity Measures for Chinese Pop Music Based on Low-level Audio Signal Attributes

In this article a method of computing similarity of two Chinese pop songs is presented. It is based on five attributes extracted from the audio signal. They include music instrument, singing voice style, singer gender, tempo, and degree of noisiness. We compare the computed similarity measures with similarity scores obtained with subjective listening by over 200 human subjects. The results show that rhythm and mood related attributes like tempo and degree of noisiness are most correlated to human perception of Chinese pop songs. Instrument and singing style are relatively less relevant. The results of subjective evaluation also indicate that the proposed method of similarity computation is fairly correlated with human perception.