Forensic Speaker Comparison by Generalized Linear Models of Articulatory and Vocal Variability in Connected Speech

Abstract

Variability is inherent to speech and arises from both speaker-related factors (e.g. sociolinguistic and personal) and linguistic factors (e.g. phonetic-phonological and coarticulatory). For the same message uttered in the same context, between-speaker variability, which can be anatomical or physiological, results from differences in vocal tract structures and motor routines, while within-speaker variability is biomechanical, stemming from variations in an individual's speech execution. Despite this understanding, the specific roles of different vocal tract components in speaker classification remain unclear, and few studies have utilized continuous speech. This study aims to model speaker variability by considering articulatory and vocal structures in continuous speech. We developed a classification procedure based on a regression model that removes part of context variability, using the residuals for speaker comparison. The development of the procedure and subsequent testing were conducted using 18 recordings from the CEFALA-1 database. Key findings include: (1) Most between speaker acoustic variability is attributed to differences related to the speaker's sex, and (2) both articulatory and vocal variables are significant for speaker classification, with vocal variables slightly outperforming articulatory variables in isolated models. Limitations of the study include its focus solely on static variability, excluding dynamic aspects, and the omission of consonant variability.