Leveraging MFCC and Chroma STFT Features for Gender and Age Group Recognition: A Comparative Machine Learning Approach

The gender and age identification of a person from his speech is a challenging task. The quality of speech can change due to a variety of factors, such as noise and unequal distribution of data. This paper examines several machine learning models for gender and age group classification with the aid of MFCC and Chroma STFT feature extraction algorithms. These algorithms effectively quantify the pitch and frequency of an audio signal. This study also handles limitations and class imbalance of data. Oversampling and augmentation techniques such as pitch shifting, noise addition, time stretching, and temporal shifting were utilized in this study. Hence, the training dataset was enriched, and the limitations of monotonous data were overcome. Further, we present a new database specifically aimed for gender and age group classification thus adding to research in this area as well. The gender recognition task achieved the best accuracy of 0.92 using SVM. For the age group recognition (3 classes) task, XGBoost overperformed the other models with an accuracy of 0.87, while ANN had 0.80 and Random Forest with 0.75. From the results, it is clear that feature fusion and data augmentation have led to enhancing speaker recognition systems. By using the findings and offered dataset, there is a possibility for further advances in utilising deep learning for actual world applications such as voice systems and demographic analysis.