Analysis of Speech Temporal Dynamics in the Context of Speaker Verification and Voice Anonymization

Abstract--In this paper, we investigate the impact of speech methods use large-scale pre-trained models for extracting specific temporal dynamics in application to automatic speaker verification attributes and provide better content and privacy preservation than and speaker voice anonymization tasks. We propose several signal processing based methods. The diversity of approaches is metrics to perform automatic speaker verification based only illustrated by the VoicePrivacy 2024 Challenge [10], which provided on phoneme durations. Experimental results demonstrate that six baseline anonymization systems, namely anonymization using x-phoneme durations leak some speaker information and can reveal vectors and a neural source-filter model [6], [11], signal processing speaker identity from both original and anonymized speech. While specific studies have been dedicated to speaker information carried by pitch [5], [6], [8], the impact of speech temporal dynamics on speaker verification and re-identification has been overlooked.