LCANC: Deep Learning-Based Long-Context Attention Framework for Adaptive Feedback Active Noise Control

In recent years, adaptive noise cancellation has developed as a vital component of modern communicative, vehicular, and industrial auditory systems. However, traditional methods significantly depend on a short-term recurrent framework, where convolutional and recurrent layers are performed separately. Hence, this research proposes a LongContext Adaptive Noise Cancellation (LCANC) model which joins a Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) with a Temporal Self-Attention mechanism (LCANC-TSA) all within one adaptive framework to allow accurate and stable noise suppression. Initially, acoustic datasets were gathered from construction, vehicle, aircraft, and industrial environments. Then, the recorded signals were preprocessed for normalization, framing, and spectral conversion to confirm better consistency of the signals. After that, the CNN layers are applied to extract local spectraltemporal features, LSTM layers is used to elucidate long-term dependencies. Also, the attention mechanism is employed for better prediction of noise without losing valuable temporal segments to be included in their anti-noise prediction. The resulting waveform is sent through a feedback structure to accurately cancel environmental noise. Experimental results demonstrated that the proposed LCANC method attains better results in terms of average noise attenuation of 26.0 dB, Root Mean Square Error (RMSE) of 0.028 compared to a Deep Learning-Based Feedback Active Noise Cancellation (DNoiseNet) model.