The Perception of Phase Intercept Distortion and its Relevance to Audio Machine Learning
Loading...
Author(s)
Vaidyanathapuram Krishnan, Venkatakrishnan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Source separation is a popular task in signal processing. It involves separating a mixture signal into its individual source signals. On the other hand, there exists an operation of applying frequency independent phase shifting. This operation distorts the waveform; this form of distortion is called phase intercept distortion. However, to humans, both the signals exactly sound the same, making it perceptually invariant. When we use source separation metrics to evaluate a phase-intercept-distorted signal as the estimated signal, it does not align well with perceptual results.
This thesis aims to solve two subproblems: (i) it proves that phase-intercept distortion is truly perceptually invariant and (ii) it improves the currently used metrics for source separation to incorporate this perceptual invariance. We have conducted experiments for each of these subproblems and reported the results. Applications of this concept in the domain of audio machine learning and source separation are also discussed, including possible directions for future work.
Sponsor
Date
2025-05-05
Extent
Resource Type
Text
Resource Subtype
Thesis