Music Structural Segmentation Across Genres with Gammatone Features

Music structural segmentation (MSS) studies to date mainly employ audio features describing the timbral, harmonic or rhythmic aspects of the music and are evaluated using datasets consisting primarily of Western music. A new dataset of Chinese traditional Jingju music with structural annotations is introduced in this paper to complement the existing evaluation framework. We discuss some statistics of the annotations analysing the inter-annotator agreements. We present two auditory features derived from the Gammatone filters based respectively on the cepstral analysis and the spectral contrast description. The Gammatone features and two commonly used features, Mel-frequency cepstral coefficients (MFCCs) and chromagram, are evaluated on the Jingju dataset as well as two existing used ones using several state-of-the-art algorithms. The investigated Gammatone features outperform MFCCs and chromagram when evaluated on the Jingju dataset and show similar performance with the Western datasets. We identify the presented Gammatone features as effective structure descriptors, especially for music lacking notable timbral or harmonic sectional variations. Results also indicate that the design of audio features and segmentation algorithms should be adapted to specific music genres to interpret individual structural patterns.