The purpose of this experiment was to analyze phonetic classification accuracy
using feature vectors composed of standard MFCCs and nonlinear invariants.  
Three invariants are tested: lyapunov exponents, correlation dimension, and 
correlation entropy.  Phone classification was performed on each frame of the 
WSJ corpus (5427149 frames total) for different combinations of features.  The 
different combinations tested were:

./exp_xx

exp_01: MFCCs + energy (baseline)
exp_02: correlation dimension
exp_03: lyapunov
exp_04: correlation entropy
exp_05: MFCCs + energy + correlation dimension
exp_06: MFCCs + energy + lyapunov
exp_07: MFCCs + energy + correlation entropy
exp_08: correlation dimension + lyapunov + correlation entropy
exp_09: MFCCs + correlation dimension + lyapunov + correlation entropy

The figures in the 'exp_xx' directory show classification counts for each 
phone.  

./compare

The figures in the 'compare' directory show a classification accuracy
comparision between 6 of the feature combinations.  From left to right, the
order is exp_01, exp_05, exp_06, exp_07, exp_08, and exp_09.  The baseline
accuracy is in the foreground of each bar for easy comparison.

./confusion_matrices

These are a first attempt at generating the normalized confusion matrices.  The lighter the color, the higher the classification count.