Euclidean Distances as measures of speaker similarity including identical twin pairs: A forensic investigation using source and filter voice characteristics. (January 2017)

• Record Type: Journal Article
• Authors: San Segundo, Eugenia; Tsanas, Athanasios; Gómez-Vilda, Pedro
• Is Part Of: Forensic science international. Volume 270(2017)
• Page Start: 25
• Page End: 38
• Publication Date: 2017-01
• Subjects: Forensic phonetics -- Acoustic analysis -- Perceptual assessment -- Voice quality -- Twins -- Pause fillers; Medical jurisprudence -- Periodicals; Chemistry, Forensic -- Periodicals; Forensic Medicine -- Periodicals; Médecine légale -- Périodiques; Chimie légale -- Périodiques; Gerechtelijke geneeskunde; Gerechtelijke chemie; Gerechtelijke psychiatrie; Chemistry, Forensic; Medical jurisprudence; Electronic journals; Periodicals; Electronic journals; 614.1
• DOI: 10.1016/j.forsciint.2016.11.020
• Languages: English
• ISSNs: 0379-0738

Abstract

Highlights: Assessment of speaker similarity combining source and filter voice characteristics. Feature selection method to determine the most parsimonious feature subset. Testing with very similar-sounding speakers, i.e. monozygotic twins (MZ). Testing using high quality and telephone-filtered recordings. Significant differences between same-speaker and different-speaker comparisons. Abstract: There is a growing consensus that hybrid approaches are necessary for successful speaker characterization in Forensic Speaker Comparison (FSC); hence this study explores the forensic potential of voice features combining source and filter characteristics. The former relate to the action of the vocal folds while the latter reflect the geometry of the speaker's vocal tract. This set of features have been extracted from pause fillers, which are long enough for robust feature estimation while spontaneous enough to be extracted from voice samples in real forensic casework. Speaker similarity was measured using standardized Euclidean Distances (ED) between pairs of speakers: 54 different-speaker (DS) comparisons, 54 same-speaker (SS) comparisons and 12 comparisons between monozygotic twins (MZ). Results revealed that the differences between DS and SS comparisons were significant in both high quality and telephone-filtered recordings, with no false rejections and limited false acceptances; this finding suggests that this set of voice features is highly speaker-dependent and therefore forensically useful. Mean ED for MZ pairs lies between the average ED for SS comparisons and DS comparisons, as expected according to the literature on twin voices. Specific cases of MZ speakers with very high ED (i.e. strong dissimilarity) are discussed in the context of sociophonetic and twin studies. A preliminary simplification of the Vocal Profile Analysis (VPA) Scheme is proposed, which enables the quantification of voice quality features in the perceptual assessment of speaker similarity, and allows for the calculation of perceptual–acoustic correlations. The adequacy of z-score normalization for this study is also discussed, as well as the relevance of heat maps for detecting the so-called phantoms in recent approaches to the biometric menagerie.