Metaheuristic Algorithms for Enhancing Multicepstral Representation in Voice Spoofing Detection: An Experimental Approach

The problem of voice spoofing detection is critical for identity authentication within biometric systems. Among the existing countermeasures, those based on soft computing have received attention from researchers in the last few years. However, it is known that spoofing representation is only effective when many features are used, which limits its applicability due to the curse of dimensionality. Accordingly, we focus on strategies to reduce the dimensionality of multicepstral features while maintaining reasonable accuracy in distinguishing between real and spoofed voices. Given the complexity of voice data, identifying and prioritizing the features with the highest information content is of utmost relevance. The study utilized four metaheuristic algorithms-GA, DA, PSO, and GWO for dimension reduction. The findings indicate that all algorithms, particularly GWO, exceed baseline performance levels. This demonstrates their efficacy in detecting voice spoofing. Moreover, it was found that certain combinations of cepstral coefficients when applied with principal component analysis projection, notably enhanced the model’s performance of voice spoofing detection.