Adaptive Comb Filtering in Speech Enhancement by Spectral Subtraction

Authors: Martin Vondra, Robert Vich

Abstract:

An improvement of the classical single-channel speech enhancement technique by comb filtering is presented in the paper. The comb filter is adaptive and tuned by the estimated pitch frequency. It is constructed empirically in the frequency domain using two approaches and is applied in voiced frames only also in frequency domain. The pitch frequency is estimated from enhanced speech in the frequency domain too. As classical speech enhancement technique, which is used prior comb filtering, the minimum mean-square error log-spectral amplitude estimator was applied. The obtained enhancement of speech corrupted by different noise sources is evaluated and compared for configuration with and without comb filtering by informal listening tests and using signal-to-noise ratio.