0
$\begingroup$

I have a novel technology that I developed, which I plan on incorporating AI beginning with simply recording a basic machine learning strategy of Input - Output for eventual training.

In this case Input will be actual musical audio (complete mastered songs).

Output is the functions of what the software does and which are correct for that particular section of the audio, etc. (This part doesn't really pertain to the question).

My question is: if I was trying to distinguish between full song audio (not necessarily a song against another song as a whole, but instead every 2-3 sec interval of a song to another), which FFT Window will yield the best data for the frequency information?

Improve this question
$\endgroup$
4
  • $\begingroup$ Wow. That might be an interesting novel technology. If you're just comparing the magnitude of the FFT of snippets of audio of equal size (those snippets are 2 or 3 seconds in length, so we're maybe 128K words, or $N=2^{17}$) , any decently tapered window will do. Say, a Kaiser window with a $\beta$ of about 6 or 7. That's a big fucking window, so you're getting a lotta frequency resolution. That may be okay, but it might give you trouble. I might consider smaller windows and a little less resolution (more smear) in the frequency domain, if I were you. $\endgroup$
    – robert bristow-johnson
    Commented Jun 16 at 2:15
  • 1
    $\begingroup$ I think it's worth asking, "How does windowing a time-domain sequence help me when performing spectral analysis? Do I really need windowing? $\endgroup$
    – Richard Lyons
    Commented Jun 16 at 2:43
  • $\begingroup$ @RichardLyons It's because I use UnityEngine as the back-end of the software and to fetch the spectral data of an audio it has to use one of the windowing functions. Not sure what other ways there are to go about it that are simple and straight-forward enough for the situation. $\endgroup$
    – Sarp Saysel
    Commented Jun 20 at 0:34
  • $\begingroup$ @robertbristow-johnson Okay - I see, thanks for your advice and input. I'll consider maybe trading off some resolution. $\endgroup$
    – Sarp Saysel
    Commented Jun 20 at 0:35

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.