A 32-point FFT based Noise Reduction Algorithm for Single Channel Speech Signals

In this paper we propose a noise reduction system for single channel speech signals. The system comprises of a 32-point FFT based spectral subtraction method, a variable least mean squares (LMS) filter and a 3-state voice activity detector. The LMS filter is designed to remove musical noise generated during spectral subtraction. It takes the original signal and the result of spectral subtraction as its inputs, and exploits the statistical differences between musical noise and speech to give a natural sounding output. The voice activity detector updates the noise spectrum estimate for spectral subtraction and controls the LMS filter parameters to make it adapt better to the incoming signal. Test cases involving three different non-stationary noise environments resulted in an average improvement of 6dB in the SNR after processing with low musical noise.