Noise Removal - any ideas???

Dear All, 

What is the best way to extract noise from data?  I have an acoustic
signal, which is fairly noisy (not sure of any of the noise parameters,
just the usual stuff you get on a microphone).  

I ask because I�m trying to tidy up the spectrogram/fft, so that it
'look's pretty' and I reckon removing the noise from the original signal
would be the way to go.  I've not dealt with real data before, and have
never thought about how best to get rid of noise.  I considered using a
threshold value but this seems rather brutish.  What are the accepted
practices for noise removal?  Or do you have any pointers on
literature/books I should be reading?

Thanks for reading!

matque

On 22 Mar, 12:01, "matque" <michelle.mcgu...@eee.strath.ac.uk> wrote:
> Dear All,
>
> What is the best way to extract noise from data?  I have an acoustic
> signal, which is fairly noisy (not sure of any of the noise parameters,
> just the usual stuff you get on a microphone).  
>
> I ask because I'm trying to tidy up the spectrogram/fft, so that it
> 'look's pretty' and I reckon removing the noise from the original signal
> would be the way to go.

Nope.

> I've not dealt with real data before, and have
> never thought about how best to get rid of noise.

The easiest noise to deal with, is the noise you never
measure. Any extra effort spent on getting a good recording
pays back with a factor 50+ in post processing.

> I considered using a
> threshold value but this seems rather brutish.  What are the accepted
> practices for noise removal?  Or do you have any pointers on
> literature/books I should be reading?

The preface of Bendat & Piersol's "Random Data" (Wiley,
2000) mentins something like (written after memory) "the
most valuable asset on any department dealing with real
life measurements is skilled, experienced personnel."

The kind of people who have seen it all befiore, done it
all before, who is able to get a system up and running
with blindfolds on.

Maybe good advice for the future, not much help right now.

The one thing you can do, is to filter the data. Apart
from that, play with the settings in the spectrogram,
like frame lengths, overlap, frequency range to display.
Different settings can give very different visual
impressions from the same data.

Rune

matque a =E9crit :
> Dear All,
>
> What is the best way to extract noise from data?  I have an acoustic
> signal, which is fairly noisy (not sure of any of the noise parameters,
> just the usual stuff you get on a microphone).
>
> I ask because I'm trying to tidy up the spectrogram/fft, so that it
> 'look's pretty' and I reckon removing the noise from the original signal
> would be the way to go.  I've not dealt with real data before, and have
> never thought about how best to get rid of noise.  I considered using a
> threshold value but this seems rather brutish.  What are the accepted
> practices for noise removal?  Or do you have any pointers on
> literature/books I should be reading?
>
> Thanks for reading!
>
> matque

Adobe Audition (formerly CoolEdit) surely can help you. Here's how it
removes noise. Basically, you record "silence" with your microphone,
then you record what you want to record. You profile the
"silence" (which really is the microphone's noise in its purest form)
supposedly in the frequency domain, and then apparently the signal you
want to filter is transformed into a frequency-time form (just like a
spectrogram) and apparently an amount of each magnitude bin is
substracted based on the pure noise's frequency profile and the signal
is transformed back to time domain. Well this explanation is based on
suppositions, but it seems to be how it works, and well you should
just go ahead and use Adobe Audtion's noise reducer.

Thanks for the suggestions!  I should have mentioned in my earlier post
that I didn�t record the data myself.  I�m working with existing data from
another university and hence have to work with these noisy signals.  This
is also the reason why I have no information on the noise, other than I
know it exists because I can see it on the plot.  

I�ve opted to just play around with the spectrogram parameters to improve
the look of the plot.  Any other suggestions?  

Matque wrote:

> I ask because I'm trying to tidy up the spectrogram/fft, so that it
> 'look's pretty' and I reckon removing the noise from the original signal
> would be the way to go.  

What don't you like about the spectrogram of your data?

On Mar 22, 9:32 am, "matque" <michelle.mcgu...@eee.strath.ac.uk>
wrote:
> Thanks for the suggestions!  I should have mentioned in my earlier post
> that I didn't record the data myself.  I'm working with existing data from
> another university and hence have to work with these noisy signals.  This
> is also the reason why I have no information on the noise, other than I
> know it exists because I can see it on the plot.  
>
> I've opted to just play around with the spectrogram parameters to improve
> the look of the plot.  Any other suggestions?  

Matque,
What you need is a non-parametric, blind adaptive scheme.  Is the
noise stationary with respect to the signal?  If so, think about
measuring the noise level between signal spurts, and develop a (S+N)/N
ratio.  Once developed, apply a gain that is inversely proportional to
the SNR.  BTW, it would work better if you decompose the signal first,
apply the scheme, then recombine.  Look up the general topic sub-band
subtraction.


Maurice Givens