How can I compare MFCC patterns using Vector Quantization Algorithm ?

I'm currently preparing a demo to defend my master thesis on an ASR system using MFCC and VQLBG .

I'm using the voice box library to compute the MFCC patterns and I use the Linde Buzo Gray Vector quantization algorithm to produce speakers codebooks . Actually each of my speakers has to produce 3 speech files for training.

Now when trying to recognize an unknown speaker, I compare his unknown codebook to each of the 3 codebooks of each well known speaker like this :

for i=1:n
    u = squareEuclidDistance(handles.unknown_speaker_vqlbg_codebook, handles.current_speaker_vqlbg_codeBook(:,:,i))
    d(1,i) = sum(min(u))/size(u,2);
end

d is a vector which should contain the minimum distance between the unknown codebook and a well known speaker codebook.

Then I keep the average of d with mean(d)) to compare it later with the average distance of all the other well known speakers. The well known speaker with the smallest average distance wins . Unfortunatly recognition rate is poor. Could anyone give me a few tips to solve this problem ?

[Htet Yamin Ko Ko]

I also used vector quantization for my research. The well known speaker with the smallest average distance also wins but check the type of the winner codebook. I set the threshold based on the experimental results and if the type of codebook is lower than the threshold, my kind of result is true. Else the results be like false. I only have to classify two types and set 1 and 2 to represent type. Sometimes the type of winner codebook is greater than 1.5 and then set it to false and Vice Versa. I hope the answer will help you.

[Htet Yamin Ko Ko]

One thing left. May be because the clusters are over-fitting. You should try some cluster numbers. I only have to classify two types but I set the required number of clusters to be 16 which gives the optimal result.