A statistical model for formant-transition microsegments of speech incorporating locus equations (Q1329435)

During the past decade an intensive research for developing some adequate framework and efficient algorithms for automatic speech recognition has been reported. Among the most successful paradigms for automatic speech recognition tasks, modeling dependencies in terms of Markov processes has proved highly desirable by its extremely consistent mathematical structure and variety of tools to allow both deep theoretical investigation and the development of practical approaches. The paper presents a locus-equations founded parametric method for modeling contextual effects in speech and the formant-transitional behaviour in consonant-vowel environment. The model uses an essentially HMM-based representation of formant-transition microsegments hence the EM algorithm can be invoked to determine the optimal model parameters. A training procedure for parameter estimation within the proposed model and the argument concerning the corresponding generalizing capacity of the consonant characteristics from a small amount of training data set are supplied in the core sections of the paper. Also, it is pointed out the value of the structural strategies in context modeling as one of the most viable means suited for unlimited vocabulary speech recognition.