Estudo e aperfeiçoamento de um vocoder de transformada senoidal

Abstract

A basic model of sinusoidal speech representation is presented and recent coding techniques of the model parameters are studied. As a complement, new methods for improving speech quality at 2.4 and 4.8 kbps using sinusoidal representation are presented. One of the methods presented is refinement in pitch estimation, since the precision of this estimation is essential for a sinusoidal harmonic model of speech. Another proposal is a more efficient procedure for the estimation of spectrum envelope parameters, which explores the psychoacoustic characteristic known as Subjective Loudness of the human hearing system. This technique was able to reduce the amount of parameters necessary to represent the envelopes of the spectrum without interfering in the final speech quality. It is also proposed an alternative approach to the synthetic phase composition step, in such a way that a better processing cost is obtained without degradation of quality. Experimental results indicate that using all of these methods in combination with the basic sinusoidal coding system increases the efficiency of the Sinusoidal Transform Vocoder operating at 2.4 and 4.8 kbps, improving the final speech quality and its Computational cost.