Within-formant spectral feature analysis for forensic speaker discrimination casework: A study of 45 Marwari monolinguals from Bikaner, India

This PhD project investigates the significance of within-formant measurements for the vowels [i:], [ɪ], [e], [ə], [a:], [o], [u:], and [ʊ], for forensic speaker comparison. It contains six traditional PhD thesis chapters providing background information, as well as three research articles presenting analyses. Data was sourced from the Marwari language, spoken in Rajasthan, India, as a testbed, but its applicability may extend to other languages. Speech was recorded from forty-five female Marwari monolingual speakers representing three caste dialects (fifteen per variety). Three speech elicitation techniques were used: reading from a wordlist, telling stories around picture stimuli, and engaging in conversation. Articles 1–3 investigate the impact of including within-formant spectral moments (i.e., centre of gravity, standard deviation, kurtosis, skewness) and spectral measures (i.e., formant amplitude, relative amplitude, spectral bandwidth, LPC bandwidth, and spectral peaks), with and without centre formant frequencies, on speaker discrimination models. The investigations encompass various combination-based systems tested against three separate variables - vowels, variety, and speech style - using linear mixed model ANOVA and linear discriminant analysis. The research contributes to existing manual systems by providing a semi-supervised feature- based system that may supplement existing ‘manual’ and semi-supervised tools. For legal systems that currently do not accept ASR analysis, it provides a more interpretable and reproducible approach.