Source Number Estimation and Clustering for Undetermined BlindSource Separation

Benedikt Loesch and Bin Yang

University of StuttgartChair of System Theory and Signal Processing

International Workshop on Acoustic Echo and Noise Control, 2008

Presenter Chia-Cheng Chen 1

Introduction

Observation Vector Clustering

Source Number Estimation

Experimental results

Conclusion

Outline

2

The task of blind source separation is to separate M

(possibly) convolutive mixtures xm[i],m = 1, . . . ,M

into N different source signals.

Present an algorithm call NOSET (Number of Source

Estimation Technique)

Introduction

3

Short Time Fourier transform (STFT)

Three steps

◦Normalization

◦Clustering

◦Reconstruction

Observation Vector Clustering(1/3)

4

Normalization

◦ The normalization is performed with respect to a reference

sensor J [4]

◦ Unit-norm normalization

Observation Vector Clustering(2/3)

5

Clustering◦K-means

Reconstruction [4]

Observation Vector Clustering(3/3)

6

The phase difference among different sensors is large enough. In the low-frequency region, this is not the case and the phase estimate is rather noisy.

Only one source is dominant at a TF point [k, l].

Source Number Estimation(1/6)

7

Selection of One-Source TF Points

Power of source n

Selection of reliable TF points

Source Number Estimation(2/6)

8

DOA Estimation◦ time delay δm for sensor m

Source Number Estimation(3/6)

9

Source Number Estimation(4/6)

10

Source Number Estimation(5/6)

11

Source Number Estimation(6/6)

12

Frequency of fs = 8 kHz and a cross-array with M = 5 microphones

16 sets of 6 speech signals (3 male, 3 female, different for each of the 16 sets)

SNR was between 20 and 30 dB

Typical values are: fl = 250Hz, t2 = 20 dB, t3 = 0.2

Experimental results(1/3)

13

Experimental results(2/3)

14

Experimental results(3/3)

15

Presented the NOSET algorithm to estimate the number of sources in blind source separation.

It relies on DOA estimation at selected one-source TF points and works in both overdetermined and underdetermined situations.

Conclusion

16