Um Sistema de Reconhecimento de Voz Contínua Dependente da Tarefa em Língua Portuguesa

Abstract

In this paper, we propose a continuous speech recognition system for automatic dial-up telephone calls in the Portuguese language. The system allows a more friendly operation, avoiding that the user he constrained to digit repetitions. The phonetic units are obtained from a reduced inventory recently proposed for the Portuguese language. A finite state machine configuration is proposed for the recognition process, which is based upon the structure of the Portuguese grammar. Using this configuration a new structure is proposed that incorporales task-dependent knowledge sources. Two algorithms widely used in the recognition step (the "level-building" and the "one-pass") are then analyzed and compared. Moreover, it is shown how to incorporate into the algorithms the knowledge from the state machine.