Analysis of Pathological Features for Spoof Detection

Deepfake speech, an advanced use of speech synthesis technology, presents a considerable challenge due to its highly realistic sound and the complexities involved in detecting it. The selection and analysis of effective features are essential for enhancing spoof detection capabilities. This study focuses on analyzing pathological features within the Myanmar Spoof Dataset. Spoofed speech in the dataset is created using five distinct techniques: vocoder methods with HiFiGAN and Parallel WaveGAN, pre-trained voice conversion via FreeVC, and GMM-based and Differential GMM-based voice conversion (GMMVC_DIFF) methods. In this paper, we perform a comparative analysis of pathological features, including Harmonics-to-Noise Ratio (HNR), six types of jitter, and seven shimmer features. Additionally, Cepstral Peak Prominence (CPP) features were assessed under both voiced and unvoiced speech conditions. The analysis demonstrates that these features show substantial variations across different spoofing techniques. Specifically, the voice conversion methods GMMVC_DIFF show notable differences in these features. These results highlight the pivotal role of these pathological features in enhancing the precision of future spoof detection systems.