Desenvolvimento de sistema de reconhecimento de fala em plataforma embarcada

Abstract

This work presents the design of an isoleted words and speak dependent automatic speech recognition system, implmented in a low-cost computer system using the general purpouse interpreted programming language Python®. In the development of this system, the signal processing stage was implemented using the end point detection technique, aiming detect the beginning and end of an utterance; the technique used in the speech signal feature extraction was Mel frequency cepstral coefficients technique. The training and recognition stages were performed through the pattern comparison approach using Dynamic Time Warping -based techniques. For the system design, firstly, tests were performed in order to find out which configurations of the signal processing and recognition stages produced the best results in relation to the accuracy ratio and average runtime per word. In these tests, the best accuracy rate obtained for a vocabulary of 25 pre-recorded words was 96.8% with an average execution time per word of less than 2 seconds. After obtaining the optimum configuration of the system, new tests were performed, simulating its use in real situation, where the accuracy rate in the best case was 88.16%, with a decoding time per word of less than 2 seconds.