Adaptive two-band spectral subtraction with multi-window spectral estimation

An improved spectral subtraction algorithm for enhancing speech corrupted by additive wideband noise is described. The artifactual noise introduced by spectral subtraction that is perceived as musical noise is 7 dB less than that introduced by the classical spectral subtraction algorithm of Berouti et al. (1979). Speech is decomposed into voiced and unvoiced sections. Since voiced speech is primarily stochastic at high frequencies, the voiced speech is high-pass filtered to extract its stochastic component. The cut-off frequency is estimated adaptively. Multi-window spectral estimation is used to estimate the spectrum of stochastically voiced and unvoiced speech, thereby reducing the spectral variance. A low-pass filter is used to extract the deterministic component of voiced speech. Its spectrum is estimated with a single window. Spectral subtraction is performed with the classical algorithm using the estimated spectra. Informal listening tests confirm that the new algorithm creates significantly less musical noise than the classical algorithm.