Host: Japan Society for Fuzzy Theory and Intelligent Informatics (SOFT)
In order to reduce background noise in noisy speech, we have investigated a noise reduction method based on a noise reconstruction system (NRS). The NRS uses a linear prediction error filter (LPEF) and a noise reconstruction filter (NRF). An input signal of a LPEF becomes a white signal. Assuming that background noise is generated by exciting a linear system with a white signal, the background noise can be reconstructed from white noise by estimating the linear system. It is a NRF based as system identification that estimates the linear system. However, in case a fixed step size for updating tap coefficients of a NRF is used, it is difficult to reduce the background noise while maintaining the high quality of enhanced speech. Therefore, a variable step size for normalized least mean square (NLMS) is proposed in this paper. In a speech section, a small step size is used so as not to estimate speech, while a large step size is used to track the background noise in a non-speech section.