Speech Pseudonymisation Assessment Using Voice Similarity Matrices
Paul-Gauthier Noé, Jean-François Bonastre, Driss Matrouf, Natalia Tomashenko, Andreas Nautsch, Nicholas Evans
Abstract
The proliferation of speech technologies and rising privacy legislation calls\nfor the development of privacy preservation solutions for speech applications.\nThese are essential since speech signals convey a wealth of rich, personal and\npotentially sensitive information. Anonymisation, the focus of the recent\nVoicePrivacy initiative, is one strategy to protect speaker identity\ninformation. Pseudonymisation solutions aim not only to mask the speaker\nidentity and preserve the linguistic content, quality and naturalness, as is\nthe goal of anonymisation, but also to preserve voice distinctiveness. Existing\nmetrics for the assessment of anonymisation are ill-suited and those for the\nassessment of pseudonymisation are completely lacking. Based upon voice\nsimilarity matrices, this paper proposes the first intuitive visualisation of\npseudonymisation performance for speech signals and two novel metrics for\nobjective assessment. They reflect the two, key pseudonymisation requirements\nof de-identification and voice distinctiveness.\n