1. 2010/12/11 Frequency Domain Blind Source Separation Based Noise Suppression to Hearing Aids (Part 2) Presenter: Cian-Bei Hong Advisor: Dr. Yeou-Jiunn Chen Date: 2010.12.1

2. 2010/12/12 Outline Introduction Purpose Literatures Review Materials & Methods Experiment Results Conclusions Future works

3. 2010/12/13 Materials & Methods Loudspeaker Preamplifier amplifier mic. M Preamplifier amplifier DAQ mic. 1 Mix signal Wireless link PC AnalysisSeparation Fig. 6. Flow of speech signals processing Source 1Source N

4. 2010/12/14 Materials & Methods Experimental devices –Condenser microphone –Preamplifier Analog Devices Inc: SSM2167 – Amplifier National Semiconductor: LM386 –Data acquisition National Instruments: WLS-9234

5. 2010/12/15 Materials & Methods STFT Centering & Whitening JADE Permutation & Scaling ISTFT Mixed Signals Separated Signals Fig. 7. Block diagram of frequency domain BSS structure

6. 2010/12/16 Materials & Methods FDICA procedure DFT ICA PCA, Centering & Whitening Solve permutation & scaling problems IDFT ( every frame )

7. 2010/12/17 Materials & Methods Centering –Subtract the mean vector Whitening –The covariance matrix of x equals the identity matrix –Using the eigenvalue decomposition (EVD)

8. 2010/12/18 Materials & Methods Joint approximate diagonalization of eigenmatrices (JADE) –Based on fourth-order cumulant tensors approach –In the ICA model the observed signals are assumed to be linear mixture of unknown variable Where H denote a unknown mixing matrix, s is the original signals –The estimated signals are defined as Where W denote a demixing matrix

9. 2010/12/19 Materials & Methods –The z is the whitening observed signals, M is a random matrix, the fourth-order cumulant matrix defined as Where k is the fourth-order cumulant, m is the element of M –Define V = UH, z = Vs, therefore Q z (M) is a symmetric matrix, to be diagonalization

10. 2010/12/110 Materials & Methods –The separated signals are obtained as –For improve the performance of separated component estimation, using multiple random matrix and the non- diagonalization level measures as

11. 2010/12/111 Materials & Methods Permutation problem –Using inter-frequency correlation-based method Fig. 8. The same source components have high correlations at neighboring frequencies

12. 2010/12/112 Materials & Methods –Define the correlation of two signals y 1 and y 2 as Where  is the mean,  is the standard deviation –Let  f be a permutation corresponding to the inverse P( f ) of the permutation matrix Where  f denote the permutation at frequency g

13. 2010/12/113 Materials & Methods Scaling problem –Let H( f ) be the unknown mixing matrix, the diagonal matrix  ( f ) should satisfy –The scaling ambiguity decided by Where P( f ) is the permutation matrix, W( f ) is the separated matrix

14. 2010/12/114 Reference [1] S. M. Lee, J. H. Won, S. Y. Kwon, Y. C. Park, I. Y. Kim, S. I. Kim, “New idea of hearing aid algorithm to enhance speech discrimination in a noisy environment and its experimental results,” International Conference of the IEEE Engineering in Medicine and Biology Society (IEMBS ’04), pp. 976-978, 2004 [2] R. Mukai, H. Sawada, S. Araki, S. Makino, “Frequency Domain Blind Source Separation of Many Speech Signals Using Near-field and Far-field Models,” EURASIP Journal on Applied Signal Processing, vol. 2006, pp. 1-13, 2006 [3] A. Hyvärinen, J. Karhunen, E. Oja, Independent Component Analysis, John Wiley & Sons, Inc, 2001 [4] J. F. Cardoso, “High-order contrasts for independent component analysis,” Neural Computation, vol. 11, pp. 157-192, 1999 [5] H. Sawada, R. Mukai, S. Araki, and S. Makino, “A robust and precise method for solving the permutation problem of frequency-domain blind source separation,” IEEE Transactions on Speech and Audio Processing, vol. 12, no. 5, pp. 530-538, 2004