Hi there!
I’m a PhD researcher in the Probabilistic Systems, Information, and Inference Group at the University of Cambridge, developing methods for statistical signal processing, machine learning, and audio and music understanding.
My research spans high-resolution time-frequency analysis, beat and metre inference, multimodal tracking, and physics-informed learning.
I will begin an EPSRC Postdoctoral Pathway Fellowship at Cambridge in 2027.
Selected Research Outputs
PHAST-Net: Attention-Guided, Physics-Informed Network for Estimation of Ideal Time-Frequency Representations
James M. Cozens, Simon J. Godsill
2026 · Preprint; submitted to IEEE Transactions on Audio, Speech and Language Processing.
Developed PHAST-Net, an attention-guided framework combining newly derived Continuous Log-frequency Adaptive Wavelet Transforms (CLAWTs) with a physics-informed reprojection objective to estimate high-resolution Ideal Time-Frequency Representations. Harmonic and spline variants enable fundamental isolation and component-wise reconstruction across spectral, tempo-based, and metrical representations. The method achieves the highest mean benchmark performance among the evaluated methods.
RIFT: Entropy–Optimised Fractional Wavelet Constellations for Ideal Time–Frequency Estimation
James M. Cozens, Simon J. Godsill
2026 · Preprint; submitted to Signal Processing (Elsevier).
Developed RIFT, a probabilistic framework combining new geometry-adaptive fractional wavelets, local entropy-weighted kernel mixture models, and positivity-constrained deconvolution with total variation regularisation for high-resolution, cross-term-suppressed time-frequency analysis. Local phase-direction estimates support component tracking and reconstruction. The method outperforms representative time-frequency methods on the reported benchmarks.
Dynamic Time Signature Recognition, Tempo Inference, and Beat Tracking Through the Metrogram Transform
James M. Cozens, Simon J. Godsill
2024 · IEEE Open Journal of Signal Processing, vol. 5, pp. 140–149.
Also presented at IEEE ICASSP 2024, Seoul, South Korea.
Introduced the Metrogram transform and a probabilistic framework for dynamic time-signature recognition, tempo inference, and beat/bar tracking in polyphonic audio. The method handles time-varying and irregular metre, retains distributions over alternative metric interpretations, and achieves competitive beat-tracking performance on the Ballroom dataset.
Bimodal Multi-Object Localisation, Siteswap Inference, and Analysis for Competitive Juggling
James M. Cozens, Simon J. Godsill
2024 · 27th International Conference on Information Fusion (FUSION), Venice, Italy, pp. 1-8.
Developed an adaptive, real-time multi-object localisation and siteswap-inference system using a bespoke bimodal machine-learning-enhanced Interacting Multiple Model (IMM) for airborne and caught-object dynamics. Enables tracking and pattern inference with up to 11 objects, alongside performance analysis and visualisation. Presented and tested an earlier prototype with performers at the 2022 European Juggling Convention.
Towards a Chromesthetic Visual Language for Generative Music Visualisation
James M. Cozens
2026 · Manuscript in preparation.
Developed a real-time audiovisual system combining neural-network analysis and procedural graphics to map harmony, timbre, rhythm, and key structure to colour, texture, and motion.
Relative Stability & Visualisation of Juggling Siteswap Sequences via Dynamical Modelling and Probabilistic Ranking Methodologies
James M. Cozens, Hugh Hunt, Colin Wright
2022 · Undergraduate dissertation.
Proposed probabilistic ranking methods and simulation-based visualisation of juggling siteswaps, drawing on dynamics, control theory, and mathematical modelling. Developed in collaboration with leading jugglers and mathematicians.
Improved Method and System
Additional Projects & Software
Realtime Polynomial Fractal Explorer
Python application using NumPy, CuPy and OpenCV for interactive GPU-accelerated fractal rendering, with 8K+ image export, supersampling and tiled zoom-flight video rendering.
Polyrhythmic Siteswaps & Music-Juggling Notation
Custom animations and hybrid notation connecting juggling siteswaps with polyrhythms and musical structure.
The Tubulator: PVC Tube Instrument
Designed and built a 32-note PVC pipe instrument, exploring acoustic resonance, tuning and instrument design.
Siteswap Math Playground
Mathematical exploration of juggling-pattern periods, object orbits and their connection to Landau’s function.
Chord Flow Visualiser
Interactive extension of the cycle of fifths showing chord connections, resolutions and movement through tonal space.
Siteswap Studio
Browser-based 2D & 3D juggling simulator combining trajectory visualisation, motion analysis, polyrhythmic timing and musical score/audio playback, with support for multi-juggler passing.