So basically, what I am attempting to do is design an algorithm that can take some amount of samples (say 10-20) that all have a common sound playing in them, but each with different noise (some with the same noise).

The noise isn't removable by normal processes (as the "noise" is actually the rest of a song), but I was wondering what process would be best for trying to isolate the original sound.

My original thought was, spectrally, to effectively take the minimum at a given frequency between two of the samples at a time, ideally leaving only the original sound. (i.e. the noise will not always be in the same place, meaning if it is zero in any of the samples, then it is not part of the original sound).

This sounds great, but due to my limited knowledge of FFTs I feel that I am underestimating or overlooking something significant (such as an FFT not being close enough to 100% accurate to the original sample for this to work).

  • Is the common signal exactly the same (i.e. digital copies) in each sample, or is it just something that sounds always the same (e.g. multiple recording-takes of the same track)? Jul 11, 2015 at 13:26
  • exactly the same. It's the same sample used multiple times. The only thing I could imagine is different is slight amplitude differences in some frequencies due to a compressor+limiter.
    – Treewhale
    Jul 12, 2015 at 2:17

1 Answer 1


You probably already thought of this, but if there is a common sound/signal frequency wise, simply taking the average of all the FFTs will attenuate the uncommon parts, while the common part remains roughly the same level.

The spectrum plot below illustrates that - in the center section you see an average of that period. The lines (the common signal) clearly stands out.

enter image description here

This might work as a way to isolate the wanted signal (e.g. by using the resulting average spectrum as a spectral mask).

  • So averaging would work? I was ultimately thinking of something like a strongly weighted average (towards the lower number). This is assuming that when you take FFTs of different sample instances that frequencies would never be missing (which I have no idea would be true or false)
    – Treewhale
    Jul 12, 2015 at 2:20
  • Yeah, I think so. I edited the answer as I was actually thinking about the FFT spectrum representations, not raw "samples". Jul 12, 2015 at 7:10
  • Sorry, I'm using "sample" to mean an instance of an audio clip, not its actual definition. In your example, is the center section obtained using the average of the two FFTs? It seems remarkably clear given the amount of other noise in each of the two.
    – Treewhale
    Jul 12, 2015 at 7:46
  • No actually, it is one complete passage where I averaged the center part. Both the top and bottom could be processed to similar smooth averages. Jul 12, 2015 at 9:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.