My recordings of bird calls have multi-harmonic noise (probably generated by some electronic device e.g. as mentioned here). I would like to perform peak-frequency and other such spectral measurements.
The multi-harmonic noise is continuous and overlaps all of my bird calls, which means any measurements will have the noise component in it. How can I 'subtract' the noise to get the true spectral measurements of the bird call itself?
I assume that the noise lines are really clipped tonal.
one way of reconstructing the signal is to use an interpolating fft (fft length > window length) so that tonals are limited to a few frequency bins. Then replace amplitude and phase by, say a linear interpolation of the adjacent (good) frequency bins. on return to time domain you should have a good approximation to the clean signal.
Edit: If you take a time bin of a (complex) spectrogram or output of FFT, then you will see the multi-harmonic lines peaking up very sharply, replacing completely the spectral information of your signal. If your signal is visible at lower and higher frequencies, then the noise, you could simply mask your interferences. this would correspond to a notch filter. you can also replace the interferences with a value that is between the upper and lower frequency bin. You do this either for amplitude and (unwraped) phase or for real and imaginary of the complex spectrum. This will not reconstruct the true signal spectrum but is the best you can do.
That's a tricky one... Is this noise somewhat constant? If so, I guess you could substract a portion of the recording without bird calls to a bird call portion. If it is not constant or if at all possible, the best solution would be to redo your recordings with a recorder and microphone that does not have electrical noise.
