PhD Research
Part 1
3D Vocal Tract Modelling from Real-time MRI
The vocal tract — the dynamic passage from the larynx to the lips — changes shape continuously during speech. Understanding its geometry is fundamental to speech science, clinical linguistics, and speaker modelling. My PhD research develops a computational pipeline for reconstructing and animating 3D models of the vocal tract from real-time MRI data.
I use constrained Large Deformation Diffeomorphic Metric Mapping (LDDMM) to perform topology-preserving morphing of vocal tract shapes across speech sounds. This is paired with deep learning segmentation networks (CNNs) for automated extraction of articulators — the tongue, lips, soft palate, and pharyngeal walls — directly from MRI frames.
The resulting 3D geometries feed into acoustic simulation pipelines (Finite Element Method acoustics via COMSOL) and 1D tube-model speech synthesisers, enabling direct validation of shape against measured formant frequencies. This bridges imaging, computational geometry, and phonetics in a single reproducible workflow.
Part 2
Acoustic–Articulatory Relationships
The second part of my PhD studies the relationship between speech acoustics and articulation, using articulatory representations derived from real-time MRI: vocal tract variables (e.g. constriction degree and location of the lips, tongue tip, tongue body and velum) and articulator contours of the tongue, lips, palate and pharyngeal walls.
I train deep learning models for acoustic-to-articulatory inversion, recovering these articulatory trajectories from the speech signal, and for articulatory speech synthesis, generating speech from articulatory representations.
The goal is to make the vocal tract visible when MRI is not available. Given only audio, the models estimate how the articulators moved, which supports speech science, clinical applications and speaker modelling without access to an MRI scanner.
Research Themes
Speech & Acoustic Analysis
Deep learning for the acoustic–articulatory relationship: acoustic-to-articulatory inversion and articulatory speech synthesis, working with vocal tract variables and MRI-derived articulators. Complemented by formant analysis, 1D tube modelling, and FEM-based acoustic simulation of MRI-derived vocal tract geometry.
Medical Imaging & Segmentation
Deep learning segmentation of anatomical structures from real-time MRI. CNN-based pipelines for delineating vocal tract articulators with high temporal and spatial resolution.
Diffeomorphic Registration
Shape analysis and morphing of 3D surfaces using LDDMM geodesic shooting. Applied to dynamic vocal tract deformation modelling across phonemes and speakers.
Machine Learning & Deep Learning
Applied ML across recommender systems, anomaly detection, and imbalanced classification. Resampling techniques (SOM-XG) for sparse data; multi-model fusion for image understanding.
Awards & Recognition
Prof. Mohan Munasinghe Award for Computer Science
Highest GPA in Computer Science — University of Colombo
Champion — International Conference in Data Science Hackathon
University of Colombo & Creative Software
Winner — Wi-Fi Based Indoor Localisation Competition
ISSAI, Nazarbayev University, Kazakhstan
Director's List — Academic Performance Award
University of Colombo