Speech enhancement using bone- and air-conducted signals and adaptive GFLANN filter

It has been widely recognized that conventional techniques and algorithms for speech enhancement indicate severe performance degradation when operated in a very harsh noise environment. In recent years, linear and nonlinear adaptive noise cancellers (ANC) have been developed for speech denoising, which use both bone- and air-conducted speech signals simultaneously to improve the enhancement quality. In this paper, we propose a nonlinear ANC which consists of a linear FIR filter and a nonlinear filter based on a generalized functional link artificial neural network (FLANN, GFLANN). Both filters are equipped in a parallel form. The proposed ANC is applied to real bone- and air-conducted speech measurements. It is revealed by extensive simulations that the proposed ANC is capable of recovering the high-frequency components of the speech signal even in a very noisy situation, and outperforms its counterparts that use the FIR filter, Volterra filter, and FLANN.