0
$\begingroup$

This is my first question for this community. My connection to DSP comes via machine learning/deep learning, and I am working in the Python ecosystem.

As a preprocessing step to audio classification, I am looking for a simple method to reduce noise.

Promising noise reduction methods such as 1 expect a noise-only sample as input, then are able to subtract noise. (The approach is also described in an answer of 2)

Getting a noise sample manually is doable for denoising a few recordings. However, it does not scale to processing thousands of recordings with different noise types and volumes.

Can you recommend a method for automatically obtaining a noise sample from every recording? Alternatively, denoising methods that do not require a noise sample as input?

$\endgroup$
2
  • $\begingroup$ What kind of noise are you looking to reduce? Are you looking to just ignore time-periods with noise, or actually denoise a signal as preprocessing? $\endgroup$
    – Jon Nordby
    Sep 26 '20 at 20:39
  • $\begingroup$ Btw, by training with noise - possibly synthetically introduced via Data Agumentation, many classifiers can become quite robust to noise without dedicated preprocessing $\endgroup$
    – Jon Nordby
    Sep 26 '20 at 20:39
0
$\begingroup$

The way I read it, the algorithm works by figuring out where the signal part lives in the frequency domain by using the statistics of the noise part. This means you don't need a noise-only signal for each signal you want to process as long as the statistics of the noise are the same for each signal realization.

For example, if you had 1000 signal realizations that you want to perform this noise reduction on, you would use a noise-only signal to compute the statistics of the noise. Then using those statistics, the algorithm will figure out signal versus part by comparing to a threshold (average noise power or something I would guess).

$\endgroup$
2
  • $\begingroup$ I believe that we cannot assume the noise statistics to be the same or sufficiently similar. The recordings are in many different environments, with different equipment and setup, ... and there are strong audible and visible differences in the amount of noise. Would you agree? $\endgroup$
    – clstaudt
    Jul 8 '20 at 14:17
  • 1
    $\begingroup$ @clstaudt Yeah you're right, that breaks my answer but I will leave it up since I do at least have the caveat "as long as the statistics of the noise are the same...". I'll be interested what others come up with $\endgroup$
    – Engineer
    Jul 8 '20 at 16:29

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.