Speech Age-Gender Classification Using Long Short-Term Memory

Recognizing the age and gender of a person with certainty from a media has a significant advantage. For example, the perpetrators recorded on a CCTV camera can be easily recognized, or someone used to lie about age on social media or job application applications can be easily detected. However, detecting the exact age of a person is still a tricky thing because of the quality media and the characteristics of the person who seems deceptive. In machine learning, the neural-based methods are commonly used for classification and recognition. However, age and gender classifications still produce unsatisfactory results, even age and gender classifications by speech are rarely discussed. Hence, the right approach is needed to create a good age and gender classification model. One of the solutions is using Recurrent Neural Network (RNN), which is made for sequential data like speech. In this paper, a speech age-gender classification model is developed using one of the popular RNN models called Long Short-Term Memory (LSTM). The experimental results show that the proposed model is trapped on the overfitting problem so that the accuracy of the testing set is lower than the training set. Regularization can reduce the difference between the accuracies of both training and testing sets but it cannot increase them. The data augmentation is able to slightly solve the overfitting problem.