A comparision of different classification systems for automatic singer identification
Emrah Karaman · DSpace Repository · 2009
Bu projede otomatik ?ark?c? tan?ma problemi i?in y?ntemler incelenmi? ve 15 ?ark?c? i?in bir otomatik ?ark?c? tan?ma sistemi olu?turulmu?tur. Sistem iki a?amadan olu?maktad?r. ?ncelikle sistem girdisi olarak kullan?lan ?ark? vokal k?s?m yani ?ark?c?n?n sesiyle beraber enstr?man seslerinin oldu?u ve vokal olmayan k?s?m yani sadece enstr?man seslerinin oldu?u k?s?m olarak ikiye ayr?l?r. Daha sonra ?ark?c? tan?ma ve s?n?fland?rma i?in tan?ma a?amas? ger?ekle?tirilir. Her iki a?amada da ses ?znitelikleri ??karma ve s?n?fland?rma i?lemleri uygulanmaktad?r. ?znitelik ??karma i?leminden ?nce ?ark?ya ?rnekleme d???rme, normalle?tirme, ?er?eveleme ve pencereleme gibi ?n i?lemler uygulan?r. ?znitelik olarak enerji, spektral ak?, s?f?r ge?i? oran?, mel frekans? cepstrum katsay?lar? (MFCC) ve do?rusal ?ng?r?l? cepstrum katsay?lar? (LPCC) kullan?l?r. Daha sonra bu ?znitelikler kullan?larak ?ark?c?y? belirlemek i?in destek vekt?r makineleri (SVM), ?oklu gauss kar???m modelleri (GMM) ve ?ok katmanl? alg?lay?c?larla (MLP) s?n?fland?r?c?lar olu?turulur. In this project, methods for automatic singer identification problem are investigated, and a singer identification system for 15 singers is implemented. The system consists of two parts. Firstly, the song as the input of the system is segmented into two parts: vocal part, which consists of the singers? voice and instrument sounds, and non-vocal part which consists of only instruments? sounds. Then, the identification step for modeling and classification of singer is applied. Both steps consist of the audio feature extraction and classification methods. In the beginning of feature extraction, preprocessing is applied to data such as down-sampling, normalization, pre-emphasizing, frame blocking and windowing. Energy, spectral flux, zero crossing rate, mel frequency cepstrum coefficients (MFCC) and linear prediction cepstrum coefficients (LPCC) are used for feature extraction. Then, support vector machine (SVM), gaussian mixture model (GMM) and multilayer perceptron (MLP) classifiers are constructed for classification of the singer with using all these extracted features.