This report provides an overview of speaker verification and its potential vulnerabilities, particularly with reference to the objectives within OCTAVE. The scope of the study covers the two individual modes of text-dependent and text-independent, which are considered in the formation of a complete voice biometric system for operating under unsupervised conditions. The report details the fundamental approaches involved, and the most recent developments in the field, including the state-of-the-art methods for speaker modelling and classification. Based on the outcomes of relevant experimental studies reported in the literature, the characteristics and performance of speaker verification in the aforementioned modes of operation are appropriately defined and analysed. Another important aspect of the study is that related to the vulnerability of speaker verification to spoofing. This is a rapidly growing challenge in the field, posing potentially serious threats to the reliability and hence usability of the technology in real applications. Various categories of spoofing, ranging from replay attacks to voice conversion, are analysed, and the level of vulnerability in each case is discussed. Whilst the study involves reference to the standard speaker verification approaches, a core facet of the report is that concerned with a commercial-grade speaker verification system developed by ValidSoft (a main industrial partner in OCTAVE). The study examines the efficiency and operational potential of ValidSoft’s speaker verification system for the realisation of the unsupervised access control required in the considered onsite and online scenarios. The detailed system analysis applied in this case is shown to be appropriately aligned with the requirements in the adopted application areas. This together with the analysis of vulnerability to spoofing are expected to facilitate the introduction of effective approaches for achieving appropriately high levels of reliability and robustness in unsupervised operating conditions.
Source: WP 3 Robustness in Speaker Verification
Dissemination level: Public