Mainstream multi-channel speech enhancement systems typically adopt a cascaded architecture that combines beamforming and post-filtering. In non-stationary noise environments, beamforming often suffers from degraded spatial filtering performance due to noise estimation errors, whereas deep learning-based post-filtering improves residual noise suppression but incurs high computational cost. This paper proposes a closed-loop enhancement framework that integrates minimum variance distortionless response (MVDR) beamformer with a multi-target generative adversarial network (MTGAN), achieving joint spatial-frequency optimization via a noise estimation feedback mechanism. In this framework, a dual-branch generator in MTGAN simultaneously performs post-filtering and noise estimation, while the estimated noise is dynamically fed back into the MVDR's covariance matrix update to enable iterative closed-loop optimization. Simulations results on public datasets show that the proposed noise feedback mechanism effectively improves the MVDR output performance. Compared with the existing MVDR-CRUSE system, the proposed MVDR+MTGAN system approach not only reduces model complexity (by 10.5% from 2.38×106 to 2.13×106 parameters) but also yields substantial gains across speech quality metrics, with a 6.56 dB increase in average segmental signal to noise ration and a 0.17 improvement in the overall composite overall voice quality prediction score (COVL). The proposed method provides an efficient and effective solution for multichannel speech enhancement in complex acoustic environments.
Wang et al. (Mon,) studied this question.