Hi, I am implementing an audio pattern recognition application, the pattern recognition is done using persons correlation coefficient on a set of N dimensional feature vectors. I have considered usage of PCA (Principal components analysis) for dimensionality reduction ( as done in some speech 2 text applications ). PCA take the most dominant dimensions of a collection of N dimensional vectors ( a matrix ), then, it take the most dominant Y<N dimensions and express the data only in those Y dimensions ( this is done using covariance matrix and Eigen vectors transformations ). Taking a noisy environment in mind, noise can cause the variance of one dimension ( out of the N dimensions used ) to become more dominant and hence, make the most dominant dimensions returned by PCA result different then if PCA was performed in a non noisy environment, saying that, it seems usage of PCA is trading performance ( smaller dimensionality ) for quality ( less resistant to noise ). Does my assumptions are true? Does usage of PCA with Audio analysis is sensitive to noisy environments? If so, what other algorithms can be used to reduce noise ( other then low pass filter )? Any help would be appreciated. Thanks, NTGO
Audio, PCA & noise reduction
Started by ●July 15, 2007