International Joint Conference on Rough Sets, Springer, Cham.
In this paper it is proposed to improve performance of the automatic speech recognition by using sequential three-way decisions. At first, the largest piecewise quasi-stationary segments are detected in the speech signal. Every segment is classified using the maximum a-posteriori (MAP) method implemented with the Kullback-Leibler minimum information discrimination principle. The three-way decisions are taken for each segment using the multiple comparisons and asymptotical properties of the Kullback-Leibler divergence. If the non-commitment option is chosen for any segment, it is divided into small subparts, and the decision-making is sequentially repeated by fusing the classification results for each subpart until accept or reject options are chosen or the size of each subpart becomes relatively low. Thus, each segment is associated with a hierarchy of variable-scale subparts (granules in rough set theory). In the experimental study the proposed procedure is used in speech recognition with Russian language. It was shown that our approach makes it possible to achieve high efficiency even in the presence of high level of noise in the observed utterance.