Search

Automatic Speaker Recognition (ASR) systems are designed to provide the user with statistics relating to the similarity of two or more speech samples and to the typicality of those shared features in the wider population. When an ASR system is used as part of a forensic investigation, the user must decide what counts as the appropriate 'wider population' and select a reference database accordingly. While it has generally been held that the voices populating the reference database should be similar in accent to that of the samples under consideration, the degree to which the accents should correspond has until now not been investigated empirically. We report in this article on a study in which the composition of the reference database was systematically varied in terms of accent, using corpora of samples of Standard Southern British English and of three subvarieties spoken in North-East England (Newcastle, Sunderland, Middlesbrough). [ABSTRACT FROM AUTHOR]

• Considerable independence of acoustic measures of source and filter. • Both source and filter encode speaker-specific information. • Degree of source-filter independence is speaker-specific. • Combining source and filter can produce optimal speaker recognition performance. This study examines the extent to which speaker-specific information is encoded in different features of vocal output and the relationships between those features. A range of acoustic features, grouped as source (laryngeal voice quality measures and fundamental frequency) and filter features (formants and Mel-frequency cepstral coefficients; MFCCs), were extracted from the vocalic portion of the hesitation marker um for 90 male speakers of Standard Southern British English. Little overall correlation between the sets of features was observed, suggesting no strong interdependence between source and filter in our data. Although filter features were consistently better at discriminating between same- and different-speaker pairs compared with source features, combining source and filter has the potential of producing the lowest error rates and the strongest speaker discrimination scores. Taken together, results show that source and filter provide complementary speaker-specific information. However, the extent of the improvements in speaker discrimination performance when combining source and filter varied across speakers. We explore potential explanations for this finding and discuss the implications for source-filter theory, and for applied fields such as speaker recognition and forensic speech science. [ABSTRACT FROM AUTHOR]

Across forensic speech science, the likelihood ratio (LR) is increasingly becoming accepted as the logically and legally correct framework for the expression of expert conclusions. However, there remain a number of theoretical and practical shortcomings in the procedures applied for computing LRs based on speech evidence. In this paper we review how the LR is currently applied to speaker comparison evidence and outline three specific areas which deserve further investigation: namely statistical modelling, issues relating to the relevant population and the combination of LRs from correlated parameters. We then consider future directions for confronting these issues and discuss the implications for forensic comparison evidence more generally. [ABSTRACT FROM AUTHOR]