Frequently at acoustic analysis in the systems of speech recognition the designer has to solve the problem of suitable feature parameters choice. Owing to consequent computer processing, the measurement of a small number of parameters with large information content is required. The shortcoming of the system in this sense is a mutual statistical dependence of measured parameters that often occurs. The paper describes a method of mutual statistical dependence evaluation by means of coherence coefficient. At the coefficient estimation the Parzen weights have been used. Some results of the method employed are presented.