Context-Aware Features for Singing Voice Detection in Polyphonic Music

The effectiveness of audio content analysis for music retrieval may be enhanced by the use of available metadata. In the present work, observed differences in singing style and instrumentation across genres are used to adapt acoustic features for the singing voice detection task. Timbral descriptors traditionally used to discriminate singing voice from accompanying instruments are complemented by new features representing the temporal dynamics of source pitch and timbre. A method to isolate the dominant source spectrum serves to increase the robustness of the extracted features in the context of polyphonic audio. While demonstrating the effectiveness of combining static and dynamic features, experiments on a culturally diverse music database clearly indicate the value of adapting feature sets to genre-specific acoustic characteristics. Thus commonly available metadata, such as genre, can be useful in the front-end of an MIR system.