A Post-Filtering Approach Based on Locally Linear Embedding Difference Compensation for Speech Enhancement

This paper presents a novel difference compensation postfiltering approach based on the locally linear embedding (LLE) algorithm for speech enhancement (SE). The main goal of the proposed post-filtering approach is to further suppress residual noises in SE-processed signals to attain improved speech quality and intelligibility. The proposed system can be divided into offline and online stages. In the offline stage, we prepare paired differences: the estimated difference of {SE-processed speech; noisy speech} and the ground-truth difference of {clean speech; noisy speech}. In the online stage, on the basis of estimated difference of a test utterance, we first predict the corresponding ground-truth difference based on the LLE algorithm, and then compensate the noisy speech with the predicted difference. In this study, we integrate a deep denoising autoencoder (DDAE) SE method with the proposed LLE-based difference compensation post-filtering approach. The experiment results reveal that the proposed post-filtering approach obviously enhanced the speech quality and intelligibility of the DDAE-based SE-processed speech in different noise types and signal-to-noise-ratio levels.