Automatic Singer Identification Systems Trained Using Speech Data

• Main Authors: Zhe-Ping Su, 蘇哲平
• Other Authors: Wei-Ho Tsai
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/ucd57z

碩士 === 國立臺北科技大學 === 電子工程系 === 106 === Intuitively, a singer identification (SID) system is trained using the singing voice of each singer to be identified. However, in most cases,it is difficult to obtain a singers a cappella in advance, especially for popular singers, because popular music usually contains background accompaniment during most or all vocal passages, which makes SID difficult. Still, we may have a chance to acquire a popular singers spoken data, for example, in a concert or a press conference, in which the speech usually does not contain background accompaniment, and therefore can be used directly to analyze the voice characteristics. Accordingly, this thesis focuses on studying whether an SID system established on the basis of a singers speech voice can still work when he/she sings. Our experiment results show that the SID accuracy achieved with Gaussian Mixture Modeling (GMM) on the Mel-Frequency Cepstral Coefficients (MFCC) is quite low, which only 38.67% by using speech voices. If we use i-Vector with Linear Discriminant Analysis (LDA), the SID accuracy can be improved to 64%.