New Similarity Scale to Measure the Difference in Like Patterns with Noise
Michihiro Jinnai, Satoru Tsuge, Shingo Kuroiwa, Fuji Ren · Institutional Repositories DataBase (IRDB) · 2009
A new similarity scale called the Geometric Distance, that numerically evaluates the degree of likeness between two patterns is proposed. Traditionally, the similarity scales known as the Euclidean distance and cosine similarity have been widely used to measure likeness. Traditional methods do not perform well in the presence of noise or pattern distortions. In this paper, a new mathematical model for a similarity scale is proposed which overcomes these limitations of the earlier models, while improving the overall recognition accuracy. Experiments in speech vowel recognition were carried out under various SNR levels in a variety of noisy environments. In all cases a significant improvement in recognition accuracy is demonstrated, with the improvement most pronounced in the noisiest conditions. In fact, at a SNR of 5 dB in a subway, the recognition accuracy improved from 65% to 75% and at 20 dB SNR from 98.4% to 99.6% over the MFCC method. Numerical modeling of simple patterns is used to demonstrate the principles behind the Geometric Distance.