Spectral subtraction is a popular method for the enhancement ofthe quality of speech corrupted by additive noise. Implementations ofspectral subtraction require an available estimate of the corruptingnoise. The spectrum of the noise is usually estimated during a period oftime known a priori to be speech free. This estimate is then assumed toremain stationary over the entire noisy speech signal. This paper makesuse of a standard spectral subtraction algorithm. However, the methoddoes not require a noise estimation obtained from a period of time whenspeech is known not to exist. Instead, use is made of a continuouslyrunning noise estimation algorithm to track the noise which is input tothe spectral subtraction process. As a result the method is novel inthat it: (1) does not require a known nonspeech interval from which todetermine the noise, and (2) can handle both nonwhite and slowly varying(relative to the speech) noise in an automatic way. Speech featureswhich are used to estimate the noise content during speech are thevoiced/unvoiced decision, pitch frequency estimate and the confidence ofthese features. Results show that the quality of speech degraded bynonwhite, nonstationary noise can be improved using spectral subtractionwith the proposed noise estimation algorithm
展开▼