Speaker‐independent recognition of unvoiced plosives using a convex time pattern of a posteriori probability Free

This paper describes a method of speaker‐independent recognition of unvoiced plosives. In this method, the convex characteristics of the time pattern of a posteriori probability is adopted for eliminating effects of speaker difference and coarticulation. The time pattern of a posteriori probability is more suitable than that of the distance pattern. In the first stage of the method, a posteriori probabilities for four categories (/p/, /t/, /k/, and silence) are calculated frame by frame from a five‐channel spectrum of five time frames using the Bayes theorem. In the next stage, the convex part of the time pattern of a posteriori probability is decided as an unvoiced plosive. The decision using the dynamic characteristics is more suitable for speaker‐independent recognition than that using static threshold. The recognition experiments are conducted for about 1400 samples of unvoiced plosives in 166 Japanese city words uttered by five male speakers. These experiments are carried out under the condition of automatic phoneme detection and without the knowledge of the following vowel. The recognition rate of 81% is obtained for the speaker‐dependent case and 59% for the speaker‐independent case. [Work supported by Grant‐in‐Aid for Scientific Research on Priority Areas, The Ministry of Education, Science and Culture of Japan.]