Redução de ruído em sinais de voz combinando filtro de Kalman e transformada Wavelet

Abstract

This work aims to present, combine and compare noise reduction techniques applied to voice signals contaminated by Gaussian white noise. The algorithms discussed in this proposal are the classic Kalman filter and wavelet transform. After application of the Kalman filter, the signal is decomposed by wavelet transform (WT) by applying the thresholding coefficients in the WT, as the last step has to signal reconstruction. The WT may decompose the signal into different coefficients with different depth levels, and thresholding can be performed in any of these coefficients, as well arise various combinations of the Kalman filter and WT. In search of the best combination of algorithms, and the best coefficients of discrete wavelet transform (DWT) and wavelet packet transform (WPD) is used Genetic Algorithm (GA). The AG is applied in the determination of wavelet transform coefficients to be filtered. The application of AG resulted in three methods are compared objectively using the distance Itakura-Saito and targeted SNR, and subjectively with a word recognition test performed on volunteers. Given the best algorithm, a comparison is made between it and the spectral subtraction in which there is the superiority of the method proposed in SNRs of 0 to -15dB. This paper discusses theoretical aspects of the two methods used, and displays in great detail the experiments.