I have very short signals (usually 100 samples from a 44100 hz audio recording ). These are 'bang' noises that a hammer makes on a piece of metal. I want to use these short signals as input to some analysis (and a machine learning algorithm). I am not experienced in sound processing but from what I've gathered one of the most informative features you can use from a sound recording is the MFCC. I am using python's scikit talkbox package to compute the mfcc of my signals. The problem is that this package requires the signal to be at least 300 samples long (and as i've said mine are about 100 samples long and sometimes shorter).

I wanted to know if there is a different process I should be doing in this case to calculate the MFCC or if there are other suggestions for other methods, besides the MFCC, that I should be using.


1 Answer 1


Have you tried zero-padding your signals? By adding zeroes to the front and/or end of the signals you could make them longer than 300 samples without affecting the outcome of your analysis.

Bear in mind that with very short signals your frequency-domain resolution is going to be pretty limited, and an impulsive sound such as that you describe is likely to have a very broad spectrum. The impact this has will depend on what you plan to do with your MFCCs, but essentially, there'll be fewer of them!

Using an existing package to perform the signal processing is generally a great idea, as long as you're happy to trust it! I'd certainly encourage you to make sure you understand all the processing being carried out on your data - this will affect your ability to demonstrate that your work is trustworthy.

  • $\begingroup$ Thanks a lot for the helpful answer! Its true that I need to understand all of the processing involved in the MFCC (If i had then I guess I would have known that zero padding would not affect the outcome analysis). In general i'm working on building a ML solution for a problem where I am completely unfamiliar with the domain of the data, so i'm trying to learn as much as I can, but I am hopeful that I am good enough at performance analysis and that it will be a good judge of whether my work is trustworthy. $\endgroup$ Nov 14, 2016 at 16:18
  • $\begingroup$ also if you have any other suggestions (perhaps via private message or if related to the question then here) I would greatly appreciate it! $\endgroup$ Nov 14, 2016 at 16:20
  • $\begingroup$ This website looks pretty good - and has a ML angle: practicalcryptography.com/miscellaneous/machine-learning/… $\endgroup$
    – Speedy
    Nov 14, 2016 at 16:23

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.