Comparison for Improvements of Singing Voice Detection System Based on Vocal Separation
share
Singing voice detection is the task to identify the frames which contain the singer vocal or not. It has been one of the main components in music information retrieval (MIR), which can be applicable to melody extraction, artist recognition, and music discovery in popular music. Although there are several methods which have been proposed, a more robust and more complete system is desired to improve the detection performance. In this paper, our motivation is to provide an extensive comparison in different stages of singing voice detection. Based on the analysis a novel method was proposed to build a more efficiently singing voice detection system. In the proposed system, there are main three parts. The first is a pre-process of singing voice separation to extract the vocal without the music. The improvements of several singing voice separation methods were compared to decide the best one which is integrated to singing voice detection system. And the second is a deep neural network based classifier to identify the given frames. Different deep models for classification were also compared. The last one is a post-process to filter out the anomaly frame on the prediction result of the classifier. The median filter and Hidden Markov Model (HMM) based filter as the post process were compared. Through the step by step module extension, the different methods were compared and analyzed. Finally, classification performance on two public datasets indicates that the proposed approach which based on the Long-term Recurrent Convolutional Networks (LRCN) model is a promising alternative.
READ FULL TEXT