A complex-valued multichannel speech enhancement learning algorithm for optimal tradeoff between noise reduction and speech distortion

摘要

To minimize speech distortion and residual noise, an optimal tradeoff between noise reduction and speech distortion needs to be considered. An optimal tradeoff method for single channel speech enhancement was presented by solving a real-valued constrained optimization model in a recent literature. This paper proposes a new optimal tradeoff method for multichannel speech enhancement by solving a complex-valued optimization problem subject to a residual noise constraint with the masking threshold of the clean speech. An effective complex-valued multichannel learning algorithm is developed and its convergence analysis is established completely in a complex domain. Experiment results confirm that the proposed multichannel speech enhancement algorithm outperforms several conventional algorithms in terms of both objective measures and subjective measures. (C) 2017 Elsevier B.V. All rights reserved.