SPEECH SEPARATION USING ICA
Correlation and Covariance matrix(with properties for dependent and independent vectors)
Kurtosis Hotelling
Transformation
Orthonormality
First moment of a random vector.
Correlation
vector.
between components of the
Covariance matrix is corresponding to correlation matrix
Covariance
Cross
matrix for two vectors x and y
correlation matrix
Measures the Gaussian nature of the signal
Independent
signals are more Gaussian.
Formula
is given as: (for column X1)
Maximum
for independent vectors
Y1(t) and Y2(t) are dependent on each other
Covariance
matrix is not a diagonal matrix for mixed signal vectors
Using
Hotelling transformation Y1(t) Z1(t) , Y2(t)
Column
Z2(t)
vectors of the transformation matrix is orthogonol in nature
Now, the covariance of LHS whereas the covariance of RHS
To get both sides equal, we require the columns of B orthonormal
Covariance of RHS
Covariance of LHS
CONSTRAINT 1 : Kurtosis values are maximum for independent signals. Thus, whatever equations we get for the coefficients of the independent signals, can be differentiated and principle of maxima can be applied. Thus, Lagrange’s method is used to estimate the values of B.
Using Hotelling transformation Y1(t) Z1(t) , Y2(t)
Z2(t)
Initializing matrix B such that BTB = [I]
Updating matrix B using Iterative formula b11 (t+1) = E[(b11 (t) z1i + b21 (t) z2i )3 z1i ]-3*b11 (t)
Making columns of matrix B orthonormal to each other B = B * real(inv(B'*B)^(1/2))
Repeating the above two steps ‘N’ times
Obtaining independent matrix signals by [BT ] * [Z] = [X]
Centering
Finding central moments
Whitening
Making coefficients uncorrelated
Denoising.
Telecommunication(CDMA). Reducing
noise in natural images(image processing)
Finding
hidden factors in financial data(econometrics)
A. Hyvärinen, J. Karhunen, E. Oja (2001): Independent Component Analysis, New York: Wiley, ISBN 978-0471-40540-5
Case Studies of Independent Component Analysis – Numerical Analysis of Linear Algebra - Alan Oursland, Judah De Paula, Nasim Mahmood
Pierre Comon (1994): Independent Component Analysis: a new concept?, Signal Processing, Elsevier, 36(3):287--314 (The original paper describing the concept of ICA)
FastICA toolbox version 2.5 - www.cis.hut.ft