I'm doing a Machine Learning project that incorporates many DSP elements in feature extraction. I am looking at the STFT of a piece of audio and trying to search for specific spectral features within it - for example, the exponential decay of the fundamental frequency of a kick drum - to determine what kind of sound the piece of audio is.

One thing I've looked into is finding the frequencies of all the local maxima of a DFT, unquantized. This includes being able to identify when the peak is "between" bins based on contextual information. A possible path to go down is to find all of the changes in direction within the spectrum, and then for N changes fit an $(N+1)$-degree polynomial using a regression, then finding all the zeroes of the derivative of this polynomial.

But does an algorithm for this already exist? It seems like something that might have been done before.


That sounds a lot like you shouldn't be using the DFT, but rather a parametric spectrum estimator.

For example, there's spectrum estimators that actually give you a function to evaluate at real-valued points rather than a vector of a sampled spectrum (your sentence "DFT, unquantized" doesn't make all that sense to me – the DFT is discrete, inherently, so what you describe as search algorithm only works if you do a massive DFT and then start somewhere in the and look for peaks from these points on).

My go-to example here is either

  • ESPRIT, which, given knowledge about the number of significant signals will just give you a vector of real-valued frequencies of peaks or
  • ROOT-MUSIC (dunno if that's the official term, it's what we used, based on [1]) which uses the pseudospectrum function generated by the MUSIC algorithm and finds the signals by looking for roots of a specifically crafted polynomial [1].

[1]: Barabell, Arthur J.: Improving the resolution performance of eigenstructure-based direction-finding algorithms. In: Acoustics, Speech, and Signal Processing, IEEE International Conference on ICASSP ’83., Band 8, April 1983, available online.

  • $\begingroup$ I will definitely look into those methods, but would upsampling the FFT also be sufficient, or are there problems with that method? I know you can upsample by zero-padding the time signal, and that will give me the peaks as estimated by an N-FFT but with additional resolution between bins. $\endgroup$ – NmdMystery Mar 29 '17 at 17:45
  • $\begingroup$ well aside from the additional computational load of doing the larger FFT and searching for a max in more bins, and the question of how much that'll actually increase your accuracy (which does incorporate your signal model)... $\endgroup$ – Marcus Müller Mar 29 '17 at 18:34
  • $\begingroup$ So, I think I should've spelled this out more clearly: Zero-padding doesn't actually increase resolution, in interpolates. See: dsp.stackexchange.com/questions/37927/… $\endgroup$ – Marcus Müller Mar 30 '17 at 12:58
  • $\begingroup$ Right, the goal here is just to find the maximum from the information I have so interpolation is fine. I just want to avoid discretized frequencies in my model. $\endgroup$ – NmdMystery Mar 30 '17 at 20:08
  • $\begingroup$ @NmdMystery that's contradicting the fact that you're using DFT – discrete fourier transform. If you can't have discretized frequencies, you must not use the DFT, padded or not. $\endgroup$ – Marcus Müller Mar 30 '17 at 23:18

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.