Three-dimensional vocal tract modelling for forensic phonetics applications

Research output: Contribution to conferenceAbstract


Recent work (Gully, Daffern & Murphy, 2018; Arnela et al., 2016) has demonstrated the viability of three-dimensional (3D) acoustic models of the vocal tract for obtaining detailed information about vocal acoustics. This paper presents an overview of a dynamic 3D vocal tract modelling technique and outlines some potential applications to forensic phonetics, including a forthcoming programme of research mapping simulated vocal tract parameters to the vocal profile analysis (VPA) scheme (Laver, 1991). The dynamic 3D vocal tract model The vocal tract model under study (Gully, Daffern, & Murphy, 2018) uses a digital waveguide mesh (DWM) modelling technique to simulate acoustic propagation through a 3D representation of the vocal tract, whose size and shape are obtained from magnetic resonance imaging (MRI) data of the upper airway. The shape of the airway is mapped to acoustic admittance parameters throughout the DWM, and this admittance map is varied during simulation runs to alter the effective shape of the airway and hence generate dynamic speech elements such as diphthongs, as illustrated in Figure 1 (overleaf). By reproducing the detailed 3D geometry of the vocal tract and applying a suitable source stimulus, synthetic speech is obtained which retains many of the perceptual identifying characteristics of natural speech (Gully, 2017). The model is therefore a useful tool for controlled study of the vocal tract in a number of applications including forensic phonetics. Forensic applications of vocal tract model The presence of a highly detailed, controllable vocal tract model has important implications for forensic phonetics, and may be leveraged to help address a number of issues. For example, by detailed morphoacoustic analysis such as that recently used for ear shapes (Zolfaghari at al., 2017), it may be possible to determine which specific structures in the vocal tract are most significant for speaker identification. Atypical vocal tract features – such as piercings, missing teeth or scarring – can be easily accounted for in the model to determine their potential effect on the acoustic output, despite their relative paucity in the population data. Furthermore, it is possible to generate stimuli for perceptual experiments based on degrees of articulatory, rather than acoustic, change, which may provide additional insight into speaker identification. The current work seeks to map the well-known vocal profile analysis (VPA) scheme (Laver, 1991) to parameters of the vocal tract model such as shape and stiffness. MRI data from trained phoneticians using the different VPA settings will be collected and used to inform the model. The result will be a valuable tool for forensic phoneticians with a number of applications including the development of training material for VPA users.
Original languageEnglish
Number of pages2
Publication statusPublished - 2018
Event27th Annual Conference of the International Association for Forensic Phonetics and Acoustics - Huddersfield, United Kingdom
Duration: 29 Jul 20181 Aug 2018


Conference27th Annual Conference of the International Association for Forensic Phonetics and Acoustics
Country/TerritoryUnited Kingdom

Cite this