0

I am trying to classify and detect sounds.

In the first phase we computed zero crossing rate of our sample sound database.

I have a sample database of about 99 sounds that includes sound from four distinct categories.

  1. Non Human Emergency sounds (Gunshots, blast)
  2. Non Human Non Emergency sounds (Car park, thunderstorm)
  3. Human Emergency sounds (Scream)
  4. Human Non Emergency sounds (talking and speech)

Next I are trying to model it with Gaussian Mixture Model (GMM).

My question is what value should be plotted in GMM? I thought may be mean and variance i.e. I compute mean of zero crossing rate and plotted it against variance but its not showing worthwhile results.

ArsalanKhairani
  • 11
  • 3
  • One feature, especially ZCR is definitely not enough to perform a reliable classification. Nevertheless. Once you have your GMM for every single model then you can calculate the likelihood of that observation under a given model - that's a standard stuff. – jojeck May 21 '15 at 15:30
  • Yes, I have read that ZCR is not enough and I will be going for other features like MFCC and/or LPC. However, the question is lets say is that I have calculated ZCR for all the sample point of a signal. How can I compute GMM from that? – ArsalanKhairani May 22 '15 at 02:11
  • First you should learn about GMM's, there is plenty of literature around in the internet and books. Here is not really a place to explain the Machine Learning fundamentals, but you might refer to [**this**](http://stackoverflow.com/questions/26019584/understanding-concept-of-gaussian-mixture-models?rq=1) answer. Also you didn't mentioned what language you are using. [**MATLAB**](http://uk.mathworks.com/help/stats/gmdistribution-class.html) or [**Python**](http://scikit-learn.org/stable/modules/generated/sklearn.mixture.GMM.html) have in-built routines that allow for very easy classification. – jojeck May 22 '15 at 07:51
  • I am using MATLAB. Most of the reference code I found about GMM starts with a 2 x n matrix. Then code then plots those points and from clusters from it. I am not clear as what should be the values in my matrix. – ArsalanKhairani May 22 '15 at 08:53
  • You have one feature, then you will get 1D plot of the Gaussian Mixture Model. There is nothing difficult about that - you will obtain one dimensional PDF . I don't agree that there are no examples for 1D GMM, you can find some [**here**](https://chrisjmccormick.wordpress.com/2014/08/04/gaussian-mixture-models-tutorial-and-matlab-code/). – jojeck May 22 '15 at 10:57
  • Related https://stackoverflow.com/questions/44760992/voice-pattern-similarity-searching-for-similar-audio-pattern – Nikolay Shmyrev Jun 26 '17 at 21:43

0 Answers0