Speech enhancement by selective spectral filtering

Raising the quality of noisy speech signals has met with limited success when only one signal is available for processing. Resynthesis techniques (e.g., LPC) eliminate noise but significantly distort the speech, whereas standard spectral subtraction methods often leave the enhanced speech still noisy and assume stationary additive noise. To enhance a noise speech signal, rather than subtracting a spectral estimate of the noise, complete elimination of the signal energy in frequency ranges where speech energy is weakest is suggested. Thus signal energy is only preserved at frequencies where the signal‐to‐noise ratio (SNR) is strongest. No modification is performed at frequencies of strong energy (unlike the spectral subtraction method), on the assumption that noise at such frequencies is masked by speech energy. Selection of which frequency ranges to pass is done on a basis of relative energy, spectral continuity (e.g., energy at a given frequency may change abruptly only at phoneme boundaries), and expect...