Thank in advance for the help.

I have several GB's of raw audio data that I have collected as part of a research study. This audio is divided up into individual samples approximately 30 seconds in length. Many samples, I would say approximately 20%, contain mild to significant noise in the form of microphone rustle (caused by an object moving across or touching the microphone)

What I would like to do is remove this noise. I do not care about recovering any audio covered by the noise; I just want to remove it via "cutting it out" (however, if i was able to recover the audio that would be ideal). I would like to remove as much of the noise as possible in a manner that doesn't lead to the audio skipping. In other words if I have noise at time 3 and time 20, it is sufficient to remove audio from time 3 to time 20 (obviously the time between noise occurances cannot be too large otherwise I'll end up removing all of my audio). It is thus okay to remove some good audio in order to maintain the playability of the sample. If I end up removing a certain percentage of audio due to noise, say more than 50%, I will just throw the sample out. I have many samples, I just need to make sure that the samples I analyze are clean. Obviously I cannot manually go through GB's worth of audio (well I could, but I don't have months).

What might be the best approach to take to identify such noise? I'm aware that I could use a representative noise sample and try to find matches in my samples, but I see two issues. First, I don't think there is a global representative sample of this noise. I want to make sure that I remove every instance of this noise, but I can't be too aggressive otherwise I'll end up with no audio. Second, I doubt this is the best approach. One can imagine the quality of audio of news reports given during a live broadcast in a storm; there is no microphone rustle from wind. There likely then is a standardized way to remove this kind of noise.

Edit: Computation is not an issue. I have access to a cluster and can code in just about any language.

Edit2: Spectrogram of a whole audio file with the microphone rustle. I would not say that this is representative though of one of my samples (in the terms of the noise it is very roughly representative)

enter image description here

Just the microphone rustle enter image description here

And a clipping of of the same audio sample with no microphone rustle enter image description here

  • 1
    $\begingroup$ Could you show spectrograms of a few representative samples that contain the rustle? $\endgroup$ – Olli Niemitalo Nov 17 '15 at 21:02
  • $\begingroup$ You need to make some plots (waveforms, spectrograms, etc) and identify some differences between the good and bad sounds. Spectral flatness, etc. $\endgroup$ – endolith Nov 17 '15 at 21:31
  • $\begingroup$ Spectrograms, not spectrums... $\endgroup$ – jojek Nov 17 '15 at 23:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.