Comparison of Different Normalization Techniques on Speakers’ Gender Detection

In this study, the effect of Short-time Mean and Variance Normalization (STMVN), Short-time Cepstral Mean and Scale Normalization (STMSN), Min-Max Normalization, Z-Score Normalization and Standard Deviation Normalization techniques on the classification performance was investigated in determining speakers’ gender. In the study, voice records which belongs to 192 male and 192 female speakers from TIMIT data set were used as data set. Features were extracted from Mel Frequency Cepstral Coefficients (MFCC) technique by using voice records and extracted features’ dimension was reduced to Principal Component Analysis (PCA), then normalized with different techniques. Support Vector Machine (SVM) was used as classifier. As a result of study, it was observed that, the highest accuracy in speakers’ gender estimation is obtained as %98.18 from features which were normalized with Standard Deviation Normalization technique and other normalization techniques were reduced accuracy