The VoicePersonae project aims to break down the barriers within the field of voice identity, which has been studied independently in different fields, and to (1) improve the accuracy of machine learning techniques for identity, (2) increase the security and robustness of biometrics, and (3) develop new techniques for privacy protection. We also aim to realize new technologies for voice conversion, biometric detection, and anonymization. In addition, we organize competitive research challenges to accelerate research in both voice conversion and biometric liveness detection, and anonymization and re-identification, which have opposing goals. In addition, a French-Japanese joint team consisting of the National Institute of Informatics (NII), the University of Avignon, and the Eurecom has been conducting research since 2018, aiming to (4) apply the research results to other modalities and establish technologies that both utilize and protect identities.
NII has achieved important results mainly in (1) and (4), for example, waveform generation technology by integrating signal processing and deep learning, fusion of different speech generation tasks, and deepfake detection, etc. Eurecom made an important contribution to (2) by organizing a worldwide ASVspoof challenge to evaluate the performance of liveness detection algorithms on a common database, and through detailed analysis, clarified the conditions necessary to build a highly accurate liveness detection model. Avignon University worked on (3), clarifying the concept of speaker anonymization, which has not been clearly defined so far, showing the necessary conditions that anonymization methods should meet, and proposing a set of evaluation metrics for it.