Singer based classification of song dataset using vocal signature inherent in signal

摘要

Singer based classification of song data is important in the applications like, organized archival and indexing of music data, music retrieval. In a song, singing voice is mixed with accompanying instrument signal. To extract the vocal characteristics of the singer, the effect of non-voiced part is to be minimized. In this work a simple methodology is proposed to remove the non-voiced segments and to reduce the impression of instruments from the voice-dominating signal. To extract the vocal signature, proposed features extract the variation pattern of zero crossing rate and short term energy. In broad sense, the features try to capture the range of pitch and energy over which a singer mostly operates. This is motivated by the way a human being tries to identify a singer. Finally, singer based classification is done using multi-layer perceptron network. Experiment is carried out with artist20 dataset and 63% classification accuracy is achieved. Comparison with reported works on the same dataset shows that the performance of the proposed simple methodology is better than the majority and very close to others.