Unified voice analysis: speaker recognition, age group and gender estimation using spectral features and machine learning classifiers

Abstract

Predicting speaker's personal traits from voice data has been a subject of attention in many fields such as forensic cases, automatic voice response systems, and biomedical applications. Within the scope of this study, gender and age group prediction was made with the voice data recorded from 24 volunteers. Mel-frequency cepstral coefficients (MFCC) were extracted from the audio data as hybrid time/frequency domain features, and fundamental frequencies and formants were extracted as frequency domain features. These obtained features were fused in a feature pool and age group and gender estimation studies were carried out with 4 different machine learning algorithms. According to the results obtained, the age groups of the participants could be classified with 93% accuracy and the genders with 99% accuracy with the Support Vector Machines algorithm. Also, speaker recognition task was successfully completed with 93% accuracy with the Support Vector Machines.