1

After training a neural network (NN) to tell the difference between a clean audio signal and a signal with a specific "noise", what is the mechanics that actually takes place where an unseen noise filled audio file gets "cleaned" up by the machine learning model?

Its not filtering nor subtraction of the model on the input audio file, as the frequency component of the wanted audio file seems "undisturbed" although the noise frequency component overlaps with portions of the desired audio file.

Thanks for your time and help in giving some guidance to an answer to this question.

Joe
  • 147
  • 4
  • 1
    The processing learned by neural networks is often referred to as a "black box" because we can't fully characterize it to understand it, that is, it's not "interpretable". This way, the processing you refer to is something that the network learns during its training but it's not something that we can interpret. You can check this answer for some more context on the interpretability of neural networks. – noe May 24 '23 at 13:16
  • 1
    I have added an answer based on my previous comment so that you can upvote/accept it if you see fit. – noe May 24 '23 at 16:03
  • 1
    Which neural network architecture? There are multiple options, and a detailed answer will depend on the particular architecture. – Jon Nordby May 28 '23 at 12:41
  • Hello, the LSTM and CNN architectures in particular. – Joe May 28 '23 at 14:35

1 Answers1

1

The processing learned by neural networks is often referred to as a "black box" because we can't fully characterize it to understand it, that is, it's not "interpretable".

This way, the processing you refer to is something that the network learns during its training but it's not something that we can interpret. Therefore, the answer to your question is that "we don't know" the exact characteristics of the processing done by the network due to its very black-box nature.

You can check this answer for some more context on the interpretability of neural networks.

noe
  • 26,410
  • 1
  • 46
  • 76