Speaker gender recognition by Parzen method

Authors: Sorokin V. N., Tananykin A. A.

Annotation: This paper presents a gender recognition study, which is based on vocal source area parameters and Parzen method. The vocal slit area parameters are derived by inverse filtering method. The probability – density function for each gender was estimated by Parzen method with common Gaussian kernel. In experiments we used a database of Russian digits recorded in comfortable conditions with several types of microphones. Clustering from the original database method selected 86 speakers, including 49 men and 37 women representing the diversity of voices. This database was used to estimate the dynamics parameters of the glottis. Obtained total recognition error rate for short vowels was about 2%.

Keywords: gender recognition, parzen method, kernel density estimation