0
$\begingroup$

I am working in this dataset data.mendeley.com/datasets/jwyy9np4gv/3? aiming to develop a machine learning model for classifying various diseases based on lung sounds during both expiration and inspiration. At present, I am in the preprocessing stage. The noise I marked in spectrogram imahe is a sthetoscope displacement noise and i want to delete it. My objective is to eliminate this noise, and I believe the most effective approach is to automatically detect and remove the intervals containing this noise from the audio signal. enter image description here

I'd appreciate any guidance on how to achieve this reduction in intensity. Additionally, I'm open to alternative suggestions for noise reduction in lung sound datasets.

Thanks in advance for your assistance!

Improve this question
$\endgroup$
3
  • $\begingroup$ Please edit your question for clarity on a couple of points. First, are these recordings that are made for human consumption, or will they feed algorithms that need to make determinations? Second, what are the "high intensity" regions? I assume that's coughing or something similar. Third, are they really recordings, or will you want to use this in some sort of a live machine, where delay matters? Forth, do you have a preferred effect you want to create? I.e., do you want to silence those intervals entirely, or attenuate them so they don't blow someone's ears out, or what? $\endgroup$
    – TimWescott
    Commented Jan 14 at 17:10
  • $\begingroup$ I am working in this dataset data.mendeley.com/datasets/jwyy9np4gv/3? aiming to develop a machine learning model for classifying various diseases based on lung sounds during both expiration and inspiration. At present, I am in the preprocessing stage. The noise I marked is a sthetoscope displacement noise and i want to delete it. My objective is to eliminate this noise, and I believe the most effective approach is to automatically detect and remove the intervals containing this noise from the audio signal. $\endgroup$
    – Zayo
    Commented Jan 14 at 18:40
  • $\begingroup$ This is StackExchange, which wants complete questions that do not require people to read the comments to understand. They want them to be paired with answers in a similar manner. Please edit your question to contain this necessary information. Thank you. $\endgroup$
    – TimWescott
    Commented Jan 14 at 21:20

1 Answer 1

Reset to default
1
$\begingroup$

If I were doing this for human consumption I'd find a metric that seems to detect the stethoscope displacement noise and mute it -- I'd probably mute it to the point where it was more pleasant (and hence, less tiring) to listen to, but still was obvious what it was.

For ML I'd be a lot more careful. The first thing I'd do is hit the books and see if there are canonical ways to make a ML algorithm correctly ignore parts of the input*. This would be much better than just masking the noise somehow in the preprocessor -- the reason being that if your masking ends up having a distinct signature (i.e., stretches of silence), then there would be a danger that the ML algorithm would be picking up on the pattern of stethoscope movement in addition to, or instead of, the lung sounds.

This probably would involve suspending updates to the training during any stretch of time that is affected by "bad" input -- which may just be the actual time of that bad input, but, if your neural net or whatever has memory of past inputs, would need to extend out after an event until the memory of the transient was flushed.

* I'm not a ML expert, but I could play one on TV.

Improve this answer
$\endgroup$

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.