A.Ya.Kalyuzhny, V.Yu.Semenov A method for speaker's gender identification on the basis of the Gaussian mixture modeling of voice acoustic parameters |
Acoustic bulletin, Vol. 12 ¹ 2, (2009) |
| The method for automatic speaker's gender classification has been proposed and its basic algorithmic stages have been described. The method is based on modeling of voice acoustic parameters distribution by a weighted sum of several Gaussian distributions (Gaussian mixture modeling, GMM). In doing so, every component of the GMM corresponds to a certain subset of voice acoustic parameters. The set of cepstral RASTA-PLP coefficients extended by the period of the basic tone has been selected as the vector of acoustic features. The male and female GMMs were trained by the expectation-maximization method initialized according to the K-means algorithm. The dependence of classification errors on the GMM types and their orders has been investigated. In different experiments, the proposed method has shown low probability of classification errors (from 9 to 0%). This fact allows the conclusion about minor importance of the GMM order and type in comparison with a necessity of the diverse presenting of the speakers in the training data set. |
|
KEY WORDS: speech processing, speaker’s voice , gender identification, the Gaussian mixture method, cepstral coefficents |
| TEXT LANGUAGE: Russian |