Statistical Analysis and Modification on the Dual-Channel Power-Level-Difference-Based Post-filter Estimator

The dual-channel power-level-difference-based (PLD-based) post-filter estimator has a bright prospect of application in the noise reduction technology of voice communication systems, while its theoretical performance and limits are still not well studied. For this purpose, this paper studies the statistical properties of the PLD-based post-filter estimator. By this study, we reveal the impacts of the three parameters including the coherence, the smoothing factor and the noise estimation error on the performance of the PLD-based post-filter estimator. Both theoretical results and simulation results indicate that the noise estimation error and the smoothing factor have significant impacts on the noise reduction performance of the traditional PLD-based post-filter estimator. According to these analysis results, a novel PLD-based post-filter estimator is proposed, which is based on the non-stationary noise estimation method and the adaptive smoothing power spectral density (PSD) estimation. Experimental results show that the proposed post-filter estimator could suppress more non-stationary noise components without introducing audible speech distortion. Moreover, the proposed algorithm performs better than other state-of-the-art dual-channel post-filter estimators in terms of both the segmental signal-to-noise-ratio improvement (SegSNRI) and the perceptual evaluation of speech quality (PESQ) score.