Noise reduction for speech signals by operations on the modulation frequency spectrum

Abstract

Summary:A new method is presented for enhancing the contrast between speech and background noise. It is based on modulation-frequency (MF) analysis, inspired by the auditory system. Instead of selecting or weighting frequency channels as in previous MF-based methods, here the MF spectrum is processed directly. The dominant common MF (mostly corresponding to ¢ ¡ ) is enhanced by an adaptive bandpass filter in each frequency channel, and the low-MF components are artificially reconstructed before transforming back. No explicit ¡ or noise-power measurement is required. Nonstationary sounds (detected by the width of the low-MF part) are specially treated by channel-dependent mixing with the original signal. The method is tested with two recognizers (at Univ. Oldenburg) and compared to classic methods such as Ephraim-Malah or spectral subtraction. Intelligibility measurements are planned.