On a mathematical procedure for detecting significant parameters in the classification of a statistical ensemble of phenomena and its applications

Summary The paper is concerned with a procedure for detecting the uncorrelated linear combinations of a given set of random variables. The linear transformation which is used is related to the eigenvectors of the covariance matrix for the original random variables. If the significant non correlated signals are in number less than the original ones (which have often a high degree of redundancy) the possibility arises for an application of this procedure to some branches of communication theory. The most important applications should be to vocoder for a further compression of information, for masking messages and also for building a machine which should be able to classify a statistical ensemble after suitable coding. The procedure has been tested by trying to extract significant parameters from a special coding of spoken Italian vowels in view of an improvement in the efficiency of speech recognition.