Spatially-Guided CNN Filter Modeling for Multi-Noise Localization and Active Noise Cancellation in Bedroom Environments

Abstract

Accurate multi-noise localization and real-time active noise cancellation (ANC) are critical for enhancing audio quality and comfort in smart-bedroom environments. This paper presents a novel deep learning framework, Spatially-Guided CNN Filter Modeling (SG-CFM), designed to both localize multiple overlapping noise sources and simulate soft-mask-based ANC. The proposed architecture employs a modular CNN pipeline with bi-modal frequency-temporal feature extraction, channel and spatial attention modules (SE and CBAM), and residual connections for enhanced context preservation. The network integrates an Enhanced Bi-Modal Convolution Block, Residual Temporal Squeeze Block, Dilated Temporal Convolution Block, and a Hierarchical Temporal Aggregation Block, which collectively model both local and long-range acoustic dependencies. The system is trained using a domain-specific dataset of bedroom noises, annotated with overlapping sound events such as traffic, fan noise, door movement, and background TV. The extracted MFCC features are passed through the spatially-tuned CNN, which outputs both multi-label predictions and a spectro temporal suppression mask simulating ANC. Experimental results demonstrate high multi-label classification performance with an average F1-score of 0.81, and effective ANC simulation by reducing noise energy in critical temporal segments. The model is lightweight, real-time capable, and suitable for deployment in embedded IoT devices, making it a strong candidate for next-generation smart home audio systems focused on sleep quality, acoustic privacy, and ambient intelligence.