Extracting vocal from a music file for singer recognition

Abstract

The purpose of this project is to achieve speaker recognition from an audio file consisting of vocal and background music. The scope of this paper will be on vocal and background separation, extracting the vocal signal for speaker recognition processing. Research for the paper is done on existing methods proposed by audio editing experts and digital signal processing researchers. 4 different approaches are proposed in this paper and experimentations are done on 3 of the suggested approaches using existing code or software available. Through experimentations, validation of the performance of the program to the project is done. From the result of the experiments, 2 methods seem more feasible. Voice Trap, a plug-in for VST, exhibits exceptional result for vocal extraction through centre-pan extraction. However, as it is commercial software, the code to the program is not made available. The other experiment, using blind source separation, displays relatively good results as well. It is able to perform satisfactory source separation for wav files supplied by the source of the code. In conclusion, the method of blind source separation and centre-pan extraction could be implemented in this project. With the available code for blind source separation, it is suggested that future improvements could be made to explore further into implementation.