I am doing some simple detection of voice. Say I have 10 samples of a person "A'S "voice.1,2,.......10 (10 samples of A uttering say "Hello"). Now I want to design a simple system which can differentiate between the voice from rest of the voices.
So I chose a simple method, where I take A's voice 1st sample and correlate with the rest of the his voices 2,3,...10. I can see that correlation coefficient comes to be say 100,105,.......99 respectively for the samples.
Now when I correlate the voice A's 1 sample with any other voice of a person saying "hello" then I am expected to get a lower correlation value. Say 50, 60 etc.
Now my questions are:
1. How can I create a classifier this case.Imean I can set a threshold say 98 manullay . But how to do it using a classifer, please send me any matlab code.
2. I am taking A's 1st sample as a base. Am I doing correct or it is a bad idea to select one sample and correlate with with rest.
if there is any other better solution to detect based on correlation, then please let me know.
My question is not with respect to voice detetion techniques rather about designing a classifier.
I'm assuming by correlation, you mean the Pearson coefficient, http://en.wikipedia.org/wiki/Pearson_product-moment_correlation_coefficient
If that's the case, what you're assembling is a series of dot products between your new sample vector, and your collection of old sample vectors. Say you have new vector v, and labelled voice data vectors (x_i, y_i). If you think about the dual form of the training optimization procedure, a linear classifier -- the logistic regression, or the support vector machine -- is going to also be making its decisions based on a weighted average of \sum_i \alpha_i x_i^Tv, where the precise values of \alpha will be set by the classifier's training algorithm as a function of how much it thinks the dot product of v with x_i is in terms of making new predictions. (Maybe your seventh recording of "HELLO" is garbled somehow; a good linear classifier can effectively learn to ignore that example.)
Your correlation approach is basically one option that a linear classifier (with an intercept term) is going to consider when it's performing its training procedure. Take a look at MATLAB's glmfit: http://www.mathworks.com/help/stats/glmfit.html It can help you produce these kinds of linear classifiers.
A suggestion to use naive Bayes classifier . Please see ref below:
http://www.statsoft.com/Textbook/Naive-Bayes-Classifier#Technical%20Notes
A longish simple minded approach would be to perhaps use the following steps:
1 Synchronize the starting point of n=0 (frame ref)
2, We have 10 sample of A 10, samples of B each of length say 20 This gives prior probability for any sig value at n( ie whether it belongs to A or B)
3. At each point ie at some n we get a new sample possibly from A or B. Now assume a certain height band ( fixed ) at that value and count events from A and B . This gives the likelihood for a new sig value say T
4. Then use P(A|T) = [P(A|T)/P(T) ] / P(A) ie aposteriori probability from likelihood and apriori to decide. Here T is the new sample value.
5. It does seem to get a bit wild with more people A, B , C ..and Strangers Sbut at least we are dealing with probability( range 0-1) and it is more intuitive ..
6. Possibly there could be scope for speeding up and improvising better algorithms along these lines
Good Luck
Cheers
from my point of view, it would be better if you try ti find out the features of A's voice and then use the classifier to classify different voices
All the best
- Badges
- Science topic
- Similar topics
- Correction