Development of the method of automatic determination of the speaker gender on the basis of joint evaluation of frequency moments of basic tons and formant frequencies

Abstract

The object of research is the methods of recognizing the speaker gender by means of speech signals. One of the most problematic places is insufficient knowledge of the choice of signs and decisive rules. This is necessary to increase the probability of correct recognition and noise immunity of gender recognition by voice signals in conditions of interference. It is also important to simplify the implementation of algorithms for recognizing the speaker gender. For recognition of the speaker gender, a new set of classification characteristics is selected, including the joint use of estimates of the average value of the pitch frequency, its kurtosis coefficient, estimates of the mean values of the formants and their asymmetry coefficients. In the course of the research, the method of statistical testing of the proposed algorithms on a personal computer is used. The experiments are carried out using real audio signals input from a microphone into a personal computer for both female and male representatives, and recorded as separate files. For this purpose, 10 standards of 10 words are used for each of the 5 female speakers and 5 male speakers. Based on the results of statistical tests for an algorithm involving the joint use of estimates of the mean value of the pitch frequency, its kurtosis coefficient, estimates of the mean values of the formants and their asymmetry coefficients, an average probability of correct recognition is obtained 1. With the additional action of additive noise of the Gaussian type, white noise and the ratio of the signal/noise q=20, for such algorithm the probability of correct recognition is experimentally obtained – 0.8. For the decision algorithm, which uses only estimates of the average value of the pitch frequency and its kurtosis coefficient, an average probability of correct recognition is estimated at 0.9. This indicates more noise immunity of such algorithms. In the future, the use of the obtained results not only for Russian and Ukrainian languages, but also for a number of foreign languages is supposed.