Estimation of Rapidly Time-Varying Harmonic Noise for Speech Enhancement

Robust speech enhancement relies on the estimation of stationary as well as non-stationary background noise. This contribution presents a novel approach for estimating the short-term power spectral densities (ST-PSDs) of rapidly time-varying harmonic noise as produced, e.g., by cars or motorcycles. The well-known Minimum Statistics algorithm is modified by frequency warping controlled by the fundamental frequency of the harmonic noise which is assumed to be known a priori. The resulting noise estimates are used for the enhancement of the noisy signal. A detailed description of the algorithm is given and it is shown by a thorough analysis that the new solution considerably outperforms two conventional noise ST-PSD estimation techniques.