Blind separation of speech signals based on ICA and DL

Independent component analysis aims to solve Colin Cherry Put forward
The famous ” cocktail lounge ” problem , In a room , Yes
Eight people are talking at the same time , At this time, there are eight wheat in different positions
Kefeng , Record eight time signals respectively $x(i)(t),i=1\cdot \cdot \cdot 8,$
If each source signal is independent of each other , Then each recorded signal
It is the weighted sum of the speech signals sent by eight speakers , With
$s(i)(t),i=1\cdot \cdot \cdot 8$ To represent eight original signals . This voice mix
The combination process can be described as shown in the right figure 2.1 Shown ：

It is easy to indicate that there is the following relationship between the original signal and the recorded signal
system ：


ICA Multiple objective functions can be selected ( Contrast function ) To optimize . First of all, from the perspective of information theory , Take the maximum likelihood principle as the starting point , There are maximum likelihood and minimum mutual information as objective functions ; Then another kind of objective function based on high-order correlation ———— Higher order cumulants and negative entropy ; Finally, the hypothesis testing principle is introduced into ICA Modeling , Get the likelihood ratio objective function .

Deep learning speech blind separation