Speech Deception Detection-Build A Model To Classify Whether Customer Can Pay The Loan Payment

This paper proposes traditional machine learning approaches and deep learning approaches for speech deception detection, utilizing a dataset based on the “Cheat-Game,” a turn-taking card game. The dataset consists of 10,788 labelled samples recorded in various environments, and the features extracted from the audio signals include cepstral and spectral features. Two different approaches, traditional and deep learning, were used to classify the audio signals for deception detection. However, the traditional machine learning approaches consistently outperformed the deep learning models in terms of accuracy. Thus, an ensemble model was constructed using the majority voting technique, which combined the best-performing traditional models. The experimental results demonstrate that the proposed approach improves the performance of deep learning models and achieves an accuracy of 78.29% and a F1 score of 78% for speech deception detection. The proposed approach has potential applications in domains, such as classifying whether customer can pay the loan or not.