Page 177 - Algorithm Collections for Digital Signal Processing Applications using MATLAB

166 Chapter 4

7. SPEECH SIGNAL SEPARATION AND DENOISING
USING INDEPENDENT COMPONENT ANALYSIS

Consider the two speakers talking in the meeting simultaneously with no
background noise. Two microphones are used for recording. One kept very
close to the first person. Another microphone is kept near to the second
person. The recorded signals of both the microphones can be treated as the
linear combinations of independent sources. [Two speech signals] Hence
ICA algorithm can be used to separate the two independent signals [Refer
chapter 2].
Consider the second situation in which single speaker is talking with the
background noise. Two microphones are used to record the signal. One
microphone kept near to the speaker. Another microphone kept near to the
assumed noise source. The recorded signals of both the microphones can be
treated as the linear combinations of independent sources [Speech signal +
Noise ].Similar to the above ICA algorithm can be used to separate the two
independent signals. Hence the ICA algorithm can be used to separate the
two independent signals and hence denoising is achieved.

7.1 Experiment 1

Two speech signals x1(t) and x2(t) are linearly mixed to get two mixed
signals y1(t) and y2(t) as given below.

y1(t)=0.7*x1(t) +0.3*x2(t)
y2(t)=0.3*x1(t)+0.7*x2(t)

The mixed signals are subjected to ICA algorithm. The independent
signals x1(t) and x2(t) are obtained as shown below. Note that FASTICA
toolbox is used to run the ICA algorithm.

Figure 4-15. Original speech signals