Estimating age group and gender information from speech signals

The rapid development of technology, the increase of large data technologies and the enhancement in occupation of data storage and processing methods has significantly increased the importance of speech recognition systems. The ability to determine the gender and age group of the speaker has great importance in speech-based applications. Studies considering application areas as distinguishing children voices are gaining significance in the process of detecting speech disorders such as dyslexia that occurs in children, or in improving interactive game programs for children. In addition, the studies have been done to distinguish adult voices can be utilized to access data which can characterize the characteristics of the speaker in criminal situations such as human abduction, threatening telephones and false alarms. It may serve in giving priority to elderly and child voices at police stations or hospital calls, withal it may lead to a better user-profiling the age interval of the customers. In this study, gender and age category of the speakers has been estimated based on the features extracted by various methods from the speech recording samples. The voices of 8 male and 8 female students were taken from each elementary school, secondary school and high school and university student groups. A total of 64 students' voice recordings were taken from these four groups. For the dataset, 32 Turkish words were chosen, some of which are high-word combinations in Turkish and some of which are frequently used random words. Mel-Frequency Cepstral Coefficients (MFCC) and Linear Predictive Cepstrum Coefficients (LPCC) methods, which are frequently used in the literature, have been used to extract the features from the speech samples. In addition, the MF&LP mixture model was tested by taking the elements of the two feature vectors together. Obtained feature vectors are classified using machine learning methods such as K nearest neighbors (KNN), Artificial neural networks (ANN), and support vector machines (SVM). Classification performance; for age group estimation is about 96% while for gender detection is around 94.6%.