In the music world, the oldest instrument is known as the singing voice that plays an important role in musical recordings. The singer's identity serves as a primary aid for people to organize, browse, and retrieve music recordings. In this thesis, we focus on the problem of singer identification based on the acoustic features of singing voice. An automatic singer identification system is constructed and has achieved a very high identification accuracy. This system consists of three crucial parts: singing voice detection, background music removal and pattern recognition. These parts are introduced and explored in great details in this thesis. To be specific, in terms of the singing voice detection, we firstly study a traditional method, double GMM. Then an improved method, namely single GMM, is proposed. The experimental result shows that the detection accuracy of single GMM can be achieved as high as 96.42%. In terms of the background music removal, Non-negative Matrix Factorization (NMF) and Robust Principal Component Analysis (RPCA) are demonstrated. The evaluation result shows that RPCA outperforms NMF. In terms of pattern recognition, we explore the algorithms of Support Vector Machine (SVM) and Gaussian Mixture Model (GMM). Based on the experimental results, it turns out that the prediction accuracy of GMM classifier is about 16% higher than SVM.

Library of Congress Subject Headings

Signal processing--Statistical methods; Pattern recognition systems; Singing--Data processing

Publication Date


Document Type


Student Type


Degree Name

Applied Statistics (MS)

Department, Program, or Center

School of Mathematical Sciences (COS)


Ernest Fokoue

Advisor/Committee Member

Joseph Voelkel

Advisor/Committee Member

Peter Bajorski


Physical copy available from RIT's Wallace Library at TK5102.9 .Y36 2016


RIT – Main Campus