Speaker-dependent differences in effective acoustic features for deepfake speech detection using identical English words

With the rapid advancement of Text-to-Speech (TTS) and Voice Conversion (VC) technologies, artificially generated speech has become highly natural, posing a serious threat to voice authentication systems. This study investigated effective acoustic features for synthetic speech detection by examining differences between bona fide and spoofed audio in the DeepFaketask of ASVspoof 2021. The authors analyzed acoustic features—including the fundamental frequency (f0) and the first through fourth formant frequencies (F1–F4)—of identical English words spoken by a single speaker, and classified them using a random forest classifier. The results indicated that F1 and F2 played a significant role in classification, with spoofed speech showing greater variability in vowel-related features. The analysis was then extended to five speakers to determine whether the single-speaker trend also held across multiple speakers. The important acoustic features varied across speakers, and in four out of five speakers, spoofed speech showed greater variability in standard deviation. Future work will further expand the variety of acoustic features and investigate features that are independent of speaker characteristics.