A New Noise Reduction Method Using Linear Prediction and System Identification

A technique that uses linear prediction and system identification to achieve noise reduction in a voice signal mixed with a background noise is proposed. In this method, the linear prediction error filter (LPEF) whose coefficients will converge such that the prediction error signal becomes white, removes the spectrum of the voice from a voice mixed with noise. Thereby, the prediction error signal has not crosscorrelation to the voice. The background noise can be reconstructed from the prediction error signal by using the adaptive digital filter (ADF) which is used as one for the system identification. The coefficients of the ADF will converge such that the error signal which is obtained by subtracting the output of the ADF from a voice mixed with noise, is minimized. Then, the ADF estimates only the background noise since that the input of the ADF (the prediction error signal) has not cross-correlation to the voice. Noise reduction is achieved by subtracting the output of the ADF from the voice mixed with noise.