Efficient voice signal feature selection approach based on Soergel distance for enhanced voice authentication on mobile devices

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This study introduces an efficient feature selection approach based on Soergel Distance (SD) for voice spoof detection. Wrapper methods such as Particle Swarm Optimization (PSO) require many iterations, depend on random exploration, and often retain irrelevant or ineffective features in the classification process, which increases computational cost. In contrast, the proposed SD approach acts as a filter method that systematically evaluates features by dividing them into blocks containing 10 to 70 samples per block and analyzing them across selection ratios from 10% to 70%. This process retains only an effective and highly consistent subset of features. For feature extraction, Constant-Q Cepstral Coefficients (CQCC) were employed, yielding 30 voice features as the initial input. Experimental results with standard classifiers, including Decision Tree (DT), Naïve Bayes (NB), K-Nearest Neighbor (KNN), Bootstrap Aggregating (Bagging), and Random Forest (RF), demonstrated the effectiveness of the proposed method, with KNN achieving the highest accuracy of 99.28%, an Equal Error Rate (EER) of 0.72%, and a minimum Tandem Detection Cost Function (min-tDCF) of 0.053 on the ASVspoof 2019 Logical Access (LA) dataset. In comparison with PSO, the accuracy gain ranges from 0.16% to 0.73%, while the proposed method retains only 21 features, with a higher reduction ratio of 3.3% over PSO. The efficiency of the proposed SD approach is demonstrated by improved classification performance, reduced dimensionality, and lower computational cost. This approach enables the use of lightweight ML models for low-power devices and real-time voice authentication systems.