FFR-based Listener Classification using Machine Learning Paper
This study delves into the comparative analysis of machine learning models for listener classification based on auditory biometrics. A series of experiments involving Support Vector Machines (SVM), 2D Convolutional Neural Networks (CNNs), and 1D CNNs were conducted to evaluate their efficacy in classifying listeners from Frequency Following Response (FFR) signals. Our findings suggest that SVMs, particularly with the Radial Basis Function (RBF) kernel, offer stable performance when trained on spectrograms, with the ability to benefit from increased data resolution. However, 2D CNNs did not achieve the same level of success, possibly due to the non-localized nature of spectral features within spectrograms. In contrast, 1D CNNs demonstrated superior proficiency in feature extraction from raw signals, showing that direct analysis of time-domain data can be highly effective. Particularly, 1D CNNs trained on raw signals matched the performance of their SVM counterparts, suggesting an inherent advantage in processing unaltered signal data. The study's outcomes highlight the significant potential of 1D CNNs in developing sophisticated auditory biometric systems and the need for balanced model selection in real-world applications considering both accuracy and computational efficiency.
This project has been conducted as a part of my PhD at the University of Ottawa, Canada.