The single channel automatic noise reduction (ANR) is primarily applied in communication systems used in noisy environments. The acoustically detected input signal consists of an overlay of the speech signal s(t) and additional noise n(t).
The algorithm's main functions are the estimation of the noise and its reduction from the input signal. Therefore, the different characteristics of speech and noise are considered. Thus, a basic assumption for a high qualitative noise supression is a low instationary noise signal.
The algorithm estimates the noise power spectrum and uses this information for a selective attenuation of a disturbed input signal. The estimation takes place in frequency domain considering typical noise characteristics, such as stationarity. For this purpose a combined algorithm based on minimum statistics and time recursive averaging is used. With every update of the noise estimate, a special adaptive filter used for the noise reduction is recalculated. A continuous update of the noise estimate even in speech frames allows for a fast adaption of the reduction filter.
As the filter calculation is based on an estimated and not on the real noise power spectrum, a tradeoff between noise suppression and speech quality have to be found.
Hence, the algorithm does not suit for complete elimination of noise. It rather improves the signal-to-noise-ratio to a certain degree. With a noise supression up to 25 dB a good speech quality can be achieved. A higher degree of supression is possible but might cause higher speech distortion.