Hi, I am trying to implement multi class SVM in speech recognition of isolated words. I am unable to implement the code for 5 classes. Can anyone help?
you can use LIBSVM tool for multi-class SVM. it is implemented in C but its interface is also available for Matlab and other languages.
The LIBSVM and LIBLINEAR are the popular tools:
http://www.csie.ntu.edu.tw/~cjlin/libsvm/
http://www.csie.ntu.edu.tw/~cjlin/liblinear/
you can also use WEKA software..
http://www.cs.waikato.ac.nz/ml/weka/
I agree with Brajesh -- libsvm in matlab should be easy to implement the classification part. You'll also need some simply way to decompose the acoustics of the speech into a lower dimensionality set of features that can be fed to the classifier. As a start, you might just use a set of Cepstral coefficients to represent a given short chunk of the input (maybe 50 msec) and represent a word as a matrix of such coefficients. I'm sure there is matlab code posted around for converting a chunk of acoustic input to Cepstral coefficients. Of course, you'll have an uneven size matrix for various inputs (since words have different lengths); SVM won't like that -- it expects equal size inputs. Not sure I have a clear answer for you on the best way to do the time-warping.
You can also do it with R. Packages e1071 and kernlab (both widely used) can be used for this purpose (in fact, both internally use the LIBSVM tool).
Cheers,
Javier
- Badges
- Science topic