Comb filter decomposition for robust ASR

The harmonic structure of the voiced speech is an effective way of conveying information in a way that is robust to white Gaussian additive noise. In this paper we propose Comb Filter Decomposition (CFD), a new method for approximating the magnitude of the speech spectrum in terms of its harmonics, which first leads to a new interpretation of the normalized autocorrelation function. Then we introduce some feature extraction methods based on CFD and on standard autocorrelation, that emphasize the harmonic peaks of the speech spectrum. The results show an improved ASR performance under noisy conditions.