Modelos da lingua para o portugues do Brasil aplicados ao reconhecimento de fala continua : modelos lineares e modelos hierarquicos (Parsing)

AUTOR(ES)
DATA DE PUBLICAÇÃO

1999

RESUMO

: Continuous speech recognition based only on acoustic information does not provide good results [Lippmann97]. Language Models can be used to characterize, capture and exploit the language regularities, improving the recognition system performance. This Thesis presents the study and implementation of Language Models for Brazilian Portuguese. Two word class bigram language models (linear models) and one model based on context-free grammar (hierarchical model) are proposed. One bigram model uses manual classification of words (grammatical classes) and the other one uses automatic classification performed by the Simulated Annealing algorithm. The grammar-based model is implemented employing a theory of Generative Grammar [Chomsky65] and uses the X-bar Syntax [Jackendoft77]. The Language Models are evaluated using a speaker dependent continuous speech recognition system developed by Morais [Morais97]. This work also presents a decoding algorithm, based on Level Building algorithm [Rabiner*85], which considers the Language Model constraints during the search procedure

ASSUNTO(S)

gramatica gerativa modelos linguisticos reconhecimento automatico da voz

Documentos Relacionados