spectral subtraction code explain

[akshatanayak]

Could someone please explain me this line in the code. This is used to calculate the number of silence segments in the degraded speech signal. The code is a part of boll spectral subtraction method. I have attached the whole code below.

NIS=fix((IS*fs-W)/(SP*W) +1); %number of initial silence segments

Why do we need to find the number of silence segments if we can work on the number of samples?

 

[Mityan]

I guess there is because of function 'segment'.
Look at the help:
segment builds models .... assuming that the model parameters are piecewise constant over time. It results in a model that has split the data record into segments over which the model remains constant.
Maybe there is no fundamental difference, but for estimating noise using a function 'mean' over several initial windows W just takes less computations than working on samples.

 

[gdarmon]

fix= means close to zero for example 1.9 becomes 1.0 .

NIS means the number of segments of initial noise you will assume that there are in the beginning of the signal usually 250ms are store and if u are recording in a 10khz and shiftrate of the overlapping segements of sound is 0.4= meanning 40% overlap between each segment of 256 samples you will get 23 segments of initial noise.
later on spectral subtraction will use this to do the average noise

 

[akshatanayak]

fix= means close to zero for example 1.9 becomes 1.0 .

NIS means the number of segments of initial noise you will assume that there are in the beginning of the signal usually 250ms are store and if u are recording in a 10khz and shiftrate of the overlapping segements of sound is 0.4= meanning 40% overlap between each segment of 256 samples you will get 23 segments of initial noise.
later on spectral subtraction will use this to do the average noise

Even I understand the meaning of individual terms but can you explain how the whole equation helps in finding the number of segments?

 

[gdarmon]

ok so you have defined IS as 0.25 meaning 250 ms of recording without doing anything ( no subtraction just noise)
you need to find out how many samples are there is 0.25 of the second so you multiply it in the fs= 10khz ( meanning 10000 samples of noise in a second)
from that, take the number and divide it into the number of windows u are creating (sp*W) which means = how much overlapping samples are between each window.
let's say the first window is 200 samples. so u will have samples 0 to 200 , the second window will have 80 =sp*200 so it will start from 80 to 280.
in this way u can calculate how many overlapping windows you will have in the 250 ms of noise.

when u know this number NIS = 23

u will use it in line 44 in SSBoll79.m
N=mean(Y,1:NIS)')'; %initial Noise Power Spectrum mean
which basically says : do an average Magnitude for all the overlapping segments of the initial 250 ms of noise.