A complementary low-cost method for broadband noise reduction in hearing aids for medium to high SNR levels

This work presents a complementary broadband noise reduction scheme for hearing aid applications. It is designed to attenuate uncorrelated and small-correlation-length acoustic noise with controlled speech distortion. Noisy speech signals are pre-processed by the proposed strategy before being subjected to an existing narrowband noise reduction system. The clean speech signal is estimated by a convex combination of the unprocessed speech signal and the output of a linear predictor. The convex combination coefficient is adjusted to provide noise suppression while avoiding significant unvoiced utterance distortions. The proposed method is optimized to minimize speech mean-square prediction-error. A low-cost adaptive implementation is proposed and compared to the conventional adaptive linear predictor showing an improved performance, as predicted by theory. Four different objective quality measures and subjective assessment performed by normal hearing volunteers indicate that the combined use of the proposed technique with a narrowband noise reduction system consistently improves speech quality for a range of signal to noise ratios. Low-cost digital hearing aids that make use of the conventional adaptive predictor for broadband noise reduction can be easily modified to incorporate the new proposal with a minimum amount of extra computational resources.