Estudo das transformadas wavelets para utilização em reconhecimento e classificação de comandos de voz.

Abstract

Voice-recognition technologies allow computers equipped with microphones to interpret human speech for transcription or as a method of voice command. Through this technology, it is possible to perform tasks without using certain input devices such as mouse and keyboard, keeping the operator’s hands free to perform different tasks. In this context, a project has been drawn up which aims, through the pattern techniques recognition, to develop a Wavelets Transform study in order to discover the best method to perform tobacco leaf classification through voice commands without the need of any other device for data entry. The methodology used was the voice acquisition in digital form and Haar and Daubechies Wavelet Transform application for analysis. For the classification process and patterns recognition, the Principal Component Analysis and Power Cepstrum techniques have been used. Several setbacks have occurred during the qualifying round, mainly because the compressed signal was large. Thus, the use of the Power Cepstrum mathematical method was necessary, allowing signal element number reduction and making it possible to have success in the elements classification, in this case, guided by the frequency of each phoneme analysis. It was concluded that the used Wavelet Transformed enabled patterns recognition, however, there are limitations when it comes to large signs, thus requiring a pre-processing.