A deep learning-based model for noise reduction and audio quality optimisation in music transmission signals

To address noise contamination, spectral compression, and reconstruction distortion encountered in wireless transmission, remote collaboration, and embedded audio devices, this paper proposes a deep learning-based framework for audio denoising and audio quality enhancement of music-oriented transmission signals.A dual-path convolutional network incorporating frequency-domain attention and perceptually guided composite loss is then designed to model long-term noise and transient musical details simultaneously.Distinct from speech-oriented models like DCCRN, SEGAN and U-Net, the proposed method fully leverages music-specific spectral dynamics and phase consistency for high-fidelity restoration under complex distortions.Evaluated using real and simulated noise scenarios with metrics including PESQ, STOI, SI-SNR and MOS-LQO, the model achieves a mean PESQ of 3.21 and an average SI-SNR improvement of 16.4 dB, outperforming baselines.Ablation and spectral visualisation validate the key modules.With strong adaptability, the framework is applicable to real-time remote music communication, collaborative systems and high-fidelity acoustic acquisition.