Assessing models of audio-based deception detection

The librosa package was employed for extracting features from 121 audio samples that were labeled as either truth or deceit, focusing on Mel-Frequency Cepstral Coefficients (MFCCs) and other signal processing techniques. To streamline the training process, a custom function was developed to organize and compile the relevant audio features. Gradient Boosting Classifiers and Support Vector Machines (SVM) were then used as models to classify the audio samples based on their labels. Model performance was evaluated using accuracy andthe confusion matrix to assess how well they could differentiate between truthful and deceptive samples. Preprocessing steps, suchas scaling numerical features and addressing missing data, were implemented to enhance the accuracy of the classification.