AbstractThis paper proposes a method of speech analysis using a new stochastic model. This model represents two different kinds of MA parts associated with pseudo‐periodical pulse train input and Gaussian process input. Therefore, the new model describes individually two different production models for voiced and unvoiced sound in the frequency domain and can precisely realize speech production models in comparison with an ARMA model. It has been shown from experimental results that the proposed stochastic model can extract more accurate characteristics of real speech than an ARMA mode