1
$\begingroup$

I read bunch of materials for extracting feature from audio signal and they all tell me to break signal into segments, why don't we analyze all the audio signal? I don't know what are the advantages of doing that and how wide a segment should be? I only see 256 samples per frame or 512 samples per frame... what about 1028 per frame?

$\endgroup$
2
$\begingroup$

Analyzing signals per segments, with proper windowing, is a way to cope with non-stationary in audio samples. With full-size analysis, features can get mixed. Segment-splitting is thus at play in many algorithms (mp3, shazam).

The length of window is often a matter of trade-offs, between data information and computing advantages:

  • signal sampling (window length is quite meaningless without sampling rate), with respect to the following:
  • analyzing or extracting informational content from the signal: various ranges of stationarity may exist in the data, or generally useful processing features,
  • easiness in computing: the power-of-two length you mention can be beneficial (faster algorithms like in the FFT), parallel computing, dedicated hardware, closer to real-time analysis.
$\endgroup$
5
  • $\begingroup$ thank you a lot. Is that all the advantages? $\endgroup$
    – Mrnobody
    Jul 5 '20 at 15:52
  • 1
    $\begingroup$ This covers a lot, and you can unfold the three items. For instance, in computing, you can build dedicated hardware, use parallel processing, etc. $\endgroup$ Jul 5 '20 at 16:03
  • 2
    $\begingroup$ @user51307: And of course there's the issue of quasi-real-time processing. If you have to wait for the signal to be complete before you start processing you can't do real-time processing. Think of speech enhancement: in many applications you want to do that in quasi-real-time, and not in an offline mode after you've recorded the complete signal. $\endgroup$
    – Matt L.
    Jul 5 '20 at 16:15
  • 1
    $\begingroup$ Matt L. did emphasizes a real computing issue. I've heard that what made Skype much better adopted (several times ago) was it's reduced lag, with respect to the bandwidth available at that time $\endgroup$ Jul 5 '20 at 16:22
  • 1
    $\begingroup$ Yep, and we mustn't forget that these short frames (20 ms) that we find even in less latency-bound standards like MPEG-1 Audio Layer II as used e.g. by the Digital Audio Broadcasting (DAB) standard is also owed to a time where a couple milliseconds of audio posed a serious memory buffer requirement for low-cost decoder equipment. $\endgroup$ Jul 5 '20 at 16:42

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.