SNR Estimation and Enhancement of Voiced Speech Based on Periodicity Analysis

This paper proposes an algorithm that aims at analyzing and enhancing the periodicity of voiced speech in aperiodic noise. Three types of periodicity features, including normalized autocorrelation (NAC), comb filter ratio (CFR), and combination of NAC and CFR, are tried in the algorithm. The signal is decomposed into framesubband units, and the signal-to-noise ratio (SNR) of each unit is estimated based on the periodicity feature and the uncorrelated assumption of speech and noise. Based on the estimated SNR, continuous Wiener gain or binary masking gain are calculated and applied to the units. The evaluation using two instrumental measures, the overall SNR and the PESQ quality measure, shows that the combination of NAC and CFR feature with continuous Wiener gain generally performs best on a voiced utterance corpus in white, burst, and "cocktail party" noise. In comparison to two state-of-art single-microphone noise reduction algorithms with different complexity, the proposed algorithm achieves better PESQ scores.