Speech-preserving active noise control: a deep learning approach in reverberant environments

Researcher Shuning Dai has introduced a novel deep learning approach to a classic audio engineering problem: Active Noise Control (ANC). Traditional ANC systems, based on FxLMS algorithms, struggle with non-linear, real-world acoustic environments and often cancel out desired speech along with the noise. Dai's proposed system tackles this by building an end-to-end control architecture centered on a Convolutional Recurrent Network (CRN). This design uses Long Short-Term Memory (LSTM) networks to model the temporal dynamics of sound and employs complex spectrum mapping to handle non-linear distortions, moving beyond the limitations of linear assumptions.

A key innovation is the inclusion of a specialized voice retention loss function. This guides the model to selectively suppress environmental noise while identifying and preserving the spectral characteristics of a target speaker's voice. To rigorously test the system in realistic conditions, the research used the Image Source Method (ISM) to create a high-fidelity acoustic simulation that includes challenging reverberation effects. Experimental results show the Deep ANC system achieves significantly better noise reduction than traditional methods, particularly for difficult, non-stationary noises like crowd babble. Critically, evaluations using standard metrics (PESQ for quality and STOI for intelligibility) confirm the system successfully maintains the clarity and naturalness of the preserved speech.