A Novel Text - Independent Voice based Automatic Gender Recognition System

Voice based gender and age classification can be helpful in a number of Information Technology based applications with speech interfaces. The recognition has to be independent of the text of the input speech if the application is online. In this work three different feature sets were tried for text independent gender recognition. The first set is Mel-Frequency Cepstral Coefficients (MFCC) C1 to C24. A feature relevance study was undertaken using F-ratio based analysis and the first feature set was transformed accordingly to get the novel second set which is the F-ratio based dimension reduced MFCC. The third set is the weighted version of the second one. All the three sets have performed well, especially the second and third show excellent recognition performance. An effort is made to reduce computational complexity by feature dimension reduction and optimizing the number of Gaussian components of the GMM based gender classifier. This work can be extended for age based speaker classification too. A review of recent works and the previous work of the authors on text – dependent gender recognition are briefly presented for context.