Time: 2:30pm to 3:30pm AEST
Date: Thursday, 18 April 2024
Venue:, Talanoa Room (R2.301), HC Coombs Building and Online via Zoom
Meeting ID: 884 1017 6511
Meeting Password: 823724
Speaker: Michael Carne
In forensic voice comparison (FVC) the evidence is usually quantified by a set of measurements from continuously-valued acoustic features. Typically, these are either measurements based on phonetically informed features (e.g. vowel formants) or automatic acoustic features derived from signal processing techniques (e.g. MFCC). Acoustic features capture speaker-specific characteristics associated with anatomical differences between vocal tracts. Speaker-specific characteristics are also reflected in longer term patterns of language use and are quantifiable discretely; either in terms of their presence or absence in speech, or by the frequency of their occurrence. These characteristics include things like a speakers habitual lexical and syntactic choices, discourse patterns, conversational style, patterns of speech disfluency. ‘High level’ features such as these have been previously integrated into automatic speaker recognition (ASR) systems. However, little attention has been given to these feature types in likelihood ratio-based FVC (LR-based FVC).
This thesis is concerned with how lexical information obtained from speech transcriptions can be incorporated in LR-based FVC. In doing so, the research evaluates how useful lexical information is in terms of: (1) how well it distinguishes one speaker from another and; (2) robustness to forensic conditions (limited recording duration and speaking style mismatches). It also investigates methodological issues, namely: (1) evaluation of score- vs. feature-based methods for LR computation based on categorical data; and (2) evaluation feature selection methods for addressing dimensionality problems intrinsic to speech transcription data.
The presentation briefly introduces FVC in the context of an ongoing paradigm shift in the forensic comparison sciences, and the rationale for the use of lexical features. The main part of the presentation highlights the empirical and methodological contributions the research makes, which includes key findings from experimental work. It concludes with describing limitations and areas for future investigation.
