Um software de reconhecimento de voz para português brasileiro

Abstract

This work describes a speech recognition software for Brazilian Portuguese. The main objective is to build a system for large vocabulary continuous speech recognition, able to be used in real-time applications. The concepts, characteristics and all steps needed for the construction of such systems are presented. Several resources were produced and made available: acoustic and language models, new voice and text corpora. The text corpus has been built through the extraction and automatic formatting of text from newspapers on the Internet. In addition, two voice corpora were produced, one based on audiobooks and another specifically developed to simulate real-time tests. This work also proposes the use of speaker adaptation techniques for solving the acoustic mismatch problem between speech corpora. Finally, an application programming interface is presented in order to facilitate using the open-source Julius speech decoder. Performance tests are also presented, comparing the developed system with a commercial software.