Empirical Performance Analysis of Speech Based Age Classification
Sriram Ravishankar, M. Kumar · 2021 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT) · 2021
Age forecast calculations are generally being used in voice assistants and other human-PC collaboration devices. The work presented here is on speech based age classification and comparing the performances of various models. An aggregate of 170 features is extracted from each sample and 88 features are chosen for training our models. Prominent features are chosen by utilizing PCA (Principal Component Analysis) along with RFE (Redundant Feature Elimination) techniques. Experimental analysis is done on a database called Common Voice, available in the Speech Database of Mozilla which comprises of around 5,000 speech samples from various age categories. From the forecast results, it was observed that the Multi-Layer Perceptron model is reasonably good for demonstrating this difficult assertion as it can learn non-linear and complex mathematical relations through its different neural layers, executed with appropriate activation, loss functions and optimal drop out values in different layers. An increase of around 10% in the training accuracy and about 5% in the testing scores were accomplished utilizing our MLP (Multilayer Perceptron) architecture when compared with other Regression models, Decision trees, Random forests and Gradient boosting algorithms.