Implicit neural representations for time-frequency signal processing

This dissertation presents a departure from conventional audio signal processing approaches that rely on fixed-dimensional vector representations and regularly sampled time-frequency grids. We introduce a flexible framework that models audio time-frequency representations using continuous, adaptive structures, enabling greater robustness and efficiency in modern applications. First, we propose a differentiable proxy for automatically optimizing Short-Time Fourier Transform parameters, aligning time-frequency resolution with task-specific objectives. We then reformulate time-frequency representations as Point Clouds, allowing for resolution-invariant processing and effective subsampling without sacrificing performance. Building on this foundation, we employ Implicit Neural Representations to model vectors and filters as continuous functions, thereby decoupling classical algorithms like multichannel filtering and matrix factorization from fixed parameters and sampling constraints. These contributions collectively propose a unified, parameter-agnostic view of signal processing that seamlessly integrates traditional methods with modern learning-based paradigms.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Subramani, Krishna

Contributors dc:contributor

• Smaragdis, Paris
• Kim, Minje
• Hasegawa-Johnson, Mark
• Choudhury, Romit Roy

• Time-frequency Representations
• Point Clouds
• Implicit Neural Representations
• Multichannel Filtering
• Machine Learning
• Signal Processing

Statement dc:rights

• Copyright 2025 Krishna Subramani

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129827

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129827