ANALYSIS

• The SVM classification results are very bad.

• The SVM toolkit has multiple knobs which require tweaking -- our experiments need more optimization.

• Preliminary decision tree results also give high classification error.

• The problem is in the alignments, we didn't extract the correct frames

• The features do obey theoretical trends, however. Our MFCCs are very close to those created by HTK, a state-of-the-art commercial recognizer.