Treinamento discriminativo de HMMs contínuos para reconhecimento de palavras isoladas

Abstract

Introduction: speech recognition technology in the present world, objectives of this work, dissertation structure; Hidden Markov models and the ML criterion: Hidden Markov models: HMM elements, 3 HMM problems, isolated word recognition using HMMs; Discriminant training and the MMI criterion: MMI criterion, GPD segmental algorithm: steps for implementation, re-estimation of HMM parameters according to segmental GPD algorithm, normalization of parameters, summary of discriminative training algorithm for isolated words; System implemented: introduction, databases: Inatel and DECOM-UNICAMP database, developed system: Markov models for vocabulary words, parameters, training and recognition module, numerical problems due to normalization constants; Tests and analysis of the results: initial tests, determination of the learning step, determination of the number of training times, order of presentation of the locutions, determination of the set of locutions for the discriminative training