I'm currently trying to generate a set of audio files where some audio events (e.g. dog bark) are immersed in a background audio scene (e.g. crowd noise).

I understand that if the event would have happened in the original scene, then there would also be some components due to the interaction between the sound and the environment. In other words, this should be, from my understanding, a convolutive mixture.

My question is, if I approximate the mixture by summing the two signals, thus doing an instantaneous mixture, am I losing a crucial part of the audio signal? In which circumstances is this a reasonable assumption, if any? My final objective is to train a classifier with the resulting audio.

  • $\begingroup$ "if I approximate the mixture by summing the two signals" That's... how you mix things. What else would you do to mix two signals? $\endgroup$ – endolith Mar 22 '19 at 20:04
  • $\begingroup$ From my understanding of theory, you would have to convolve the audio with the impulse response of the environment in order to obtain a real simulation. $\endgroup$ – fred_101512 Mar 28 '19 at 9:16
  • $\begingroup$ That's not mixing, that's convolution. What exactly are you trying to do? If you want to combine a dog barking with crowd noise, you just sum them together. If you want to make a dog bark sound like it's in an empty cathedral, you would convolve it with the IR of the cathedral $\endgroup$ – endolith Mar 28 '19 at 20:39
  • $\begingroup$ I would like to get a realistic approximation of, say, a dog bark happening in a crowd. Unfortunately, I do not have access to impulse responses, so I was asking under what circumstances it is realistic to approximate the convolution by just summing. $\endgroup$ – fred_101512 Apr 1 '19 at 10:37
  • $\begingroup$ Oh, so you want to simulate a dog barking in a crowd of silent people, with the bark being muffled by their clothes and bodies? That would require convolving with an impulse response, yes. There are no circumstances in which it is realistic to approximate a convolution by summing. That makes no sense. They are completely different operations. $\endgroup$ – endolith Apr 1 '19 at 14:58

We typically model audio to be a pretty linear thing, so yes, just add – you can multiply each input with a constant to weigh them differently.


Your Answer

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

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