Publication

Abstract : Voice-based gender classification has potential significance in the areas of forensic, security and surveillance applications. Biometrics and computer vision [1] are the two areas where gender classification plays vital role. If the audio or the speech sample is noise free, then the recognition accuracy will be high. But, in the case of noisy audio samples and real-time data, existing algorithms failed to achieve greater accuracy. To improve the real-time gender identification, this paper introduces a novel approach that uses feature engineering with optimized Mel Frequency Cepstral Coefficients (MFCC) Features. VoxCeleb2 dataset was used for experimentation which has 6000 data samples. The features are extracted from the audio samples and optimized with polynomial features and logarithmic transformation. In addition, Optuna is used for hyperparameter optimization to improve the model’s performance. The proposed methodology was tested with multiple classifiers. An increase of 7.15% was observed on incorporating feature engineering.