Vector compression of bird songs spectra in water sites by using the linear prediction method and its application to an automated Bayesian species classification

Abstract

We propose an automated vector compression of bird songs spectra in water-sites by linear prediction method and principal component analysis, and present the experimental validation of the special features and effectiveness of the proposed method through automatic Bayesian discrimination of compressed vectors of the 14 different songs gathered. The result shows that the whole information to classify the songs completely can be fundamentally compressed to a vector of at least 30 dimensions consisting of the variance of residue series and optimal prediction coefficients. Further compression is proposed by unification of covariance matrices of different characteristics among the classes, and principal component analysis has the special features that it makes Bayesian discrimination possible even for cases where the conventional can not conduct the discrimination, with enhanced rates of 1.58 to 1.73 times and within reduction of mean correct classification rate only of 0.8%.