The design and testing of a noise reduction algorithm based on spectral subtraction

Three noise reduction algorithms based on amplitude subtraction were designed and used to process speech mixed with babble noise in two signal-to-noise ratios. The estimation of the noise-magnitude spectrum was performed with a novel synchro method, which exploits specific characteristics of the speech signal. The unprocessed and processed signals were evaluated psychoacoustically by means of a four-alternative-forced choice test with monosyllabic words (minimal pairs) in carrier phrases. The testing was carried out on groups of normally hearing and hearing-impaired subjects and the long-term power spectra of the processed signals were shaped to be essentially identical with those from the corresponding unprocessed signals. For the hearing-impaired subjects all signals were spectrally shaped according to the POGO-fitting rule. None of the algorithms improved speech intelligibility for any group of listeners and no change in the overall pattern of confusion was observed.