Voice Privacy Through Time-Scale and Pitch Modification

Abstract

An attacker can fraudulently get access (instead of the genuine user) if the users’ speech data has not been preserved by using any protection. Hence, it is important to protect users’ speech data for which a voice privacy system can be employed. A voice privacy system is not designed based on any particular kind of attack. Instead, it is designed in a generalized way, making it as universal system. This study presents the time-scale and pitch modification-based anonymization methods to modify the speaker-dependent speech parameters (i.e., $F_{0}$) for better privacy preservation of speech data. The proposed voice privacy performance is compared with the signal processing-based baseline system of the INTERSPEECH 2020 voice privacy challenge. The authors have used various perturbation methods, concluding that speed perturbation with factor 0.8 is better to get adequate speaker anonymization (with $38.5\%$ Equal Error Rate (EER) and $91.3\%$ De-IDentification (DeID)) and acceptable speech intelligibility ($4.86\%$ WER) for female speakers. It is observed that speed and pitch perturbation are two important candidates for anonymization. However, the tempo perturbation is not found to be so useful for speaker anonymization.