Reconhecimento automático de palavras isoladas, independente de locutor, para sistemas embarcados.

Abstract

In this dissertation, a speaker-independent speech recognition system for isolated words is presented. Here, we focus on battery-dependent embedded systems that have the requirement of lower power consumption. Considering this requirement and targeting a hardware implementation, we chose to use simpler techniques for recognition such as cepstrum coefficients obtained from LPC coefficients to compose the feature vector. It was also used quantization vectors to generate patterns and decision rule based on Euclidean distance. The proposed system was implemented in software and validated with a database composed of 1,232 training sentences and 770 testing sentences. It was achieved a recognition rate of 96.36%. Compared to more complex modeling that achieved recognition rate of 100% using continuous densities Hidden Markov Models, linguistic models and mel-frequency cepstrum coefficients (MFCC), the proposed technique is highly satisfactory. A significant reduction in complexity and, consequently, in power consumption, necessary for hardware implementation, is achieved by paying the price of only a small reduction in performance.