Melhoria da convergência do método ICA-MAP para remoção de ruído em sinal de voz

Abstract

Source separation consists on recovering a latent signal from a set of observable mixtures. In denoising problems, that could be faced as a source separation problem, it is necessary to extract an unobserved signal of voice contamined by noise. In such a case, an important approach is based on independent component analysis (ICA models). Especially, the use of ICA with the maximum a posteriori (MAP) estimation is known as ICA-MAP and, unlike other approaches, employs two different transformations, one for voice signal and another for noise, providing a better estimate within a linear environment. This work presents a modification to the algorithm ICA-MAP in order to improve its convergence. It has been observed through tests that it is possible to limit the magnitude of the gradient vector, used for estimating the parameters of the denoising model, and thus improving the stability of the algorithm. Such adaptation can be understood as a restriction of the original optimization problem. Another proposed approach is to approximate the derivative of the model GGM (generalized gaussiam model) around zero by using a spline function. In order to accelerate the algorithm, it is applied a variable step in the ascent gradient algorithm. Comparative tests were conducted employing standardized voice (male and female) and noise databases. At the end, the results are compared with classical techniques in order to highlight the advantages of the method.