Avaliação de redes neurais auto-organizáveis para reconhecimento de voz em sistemas embarcados

Abstract

This dissertation presents a comprehensive performance comparison study of self-organizing neural network architectures applied to speech recognition in embedded systems. In particular, we assess the feasibility of using the self-organizing map (SOM) and the tree-structured self- organizing map (TS-SOM) in the task of interest. Speech signals are parameterized through LPC and Cepstral coefficients, and both, speaker dependent and speaker independent tests, are carried out. The speech data used in this work consisted of 17 classes of words, naturally pronounced, corresponding to digits and operators required to develop a voice-driven calculator. The application is then embedded into the N95 Nokia smartphone. The obtained results indicate that the SOM and TS-SOM networks can be successfully applied to the task of interest, since they perform similarly to or even better than classical speech recognition methods, concerning recognition rates and the computational costs.