Dimension Reduction Analysis of Signal Manifold for Vowels in Time and Frequency Domain

Abstract

In this paper, the LLE and ISOMAP algorithms in manifold learning are applied them to the analysis of vowel signals in time and frequency domain. Time domain simulation results show that the two dimensionality reduction methods can implement two-dimensional visualization of signals while preserving the high-dimensional manifold structure of original signals as much as possible. The time-frequency domain dimension reduction analysis of vowel signal manifold effectively solves the problem that high-dimensional speech signals can’t be intuitively felt, and provides a new potential way for signal classification. The frequency domain analysis is further optimized on the basis of time domain simulation. Because half of the amplitude values in DFT is used in the simulation, the two-dimensional manifold of the signal is roughly linearly distributed, which can effectively reduce redundancy and make the signal more compactly expressed in the frequency domain