Noise Reduction for Wideband Speech Exploiting Spectral Dependencies Based on Conditional Estimation

Abstract:
This contribution presents a wideband (50 Hz - 7 kHz) speech enhancement system that is operating in the frequency domain. A conventional noise suppression is used in the low band, while a joint noise suppression approach is applied in the high band using techniques known from artificial bandwidth extension (BWE). Statistical dependencies between the low band (50 Hz - 4 kHz) and the high band (4 - 7 kHz) are exploited. Features from the processed (enhanced) low band signal are extracted and used to estimate subband energies of the high band. The weighting gains determined from these energy estimates are adaptively combined with conventional gains obtained in addition for the high band. As novel feature compared to "T. Esch, F. Heese, B. Geiser, and P. Vary, Wideband noise suppression supported by artificial bandwidth extension techniques", the BWE training process incorporates noisy speech and more appropriate features resulting in a better noise estimate. The performance of the proposed method is shown to be consistently better than the conventional approach and "T. Esch, F. Heese, B. Geiser, and P. Vary, Wideband noise suppression supported by artificial bandwidth extension techniques".