Article: Computer music denoising and enhancement using dual-branch communication with spectral subtraction Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.54 pp.1 - 23 Abstract: This paper proposes a deep learning-based model for computer music denoising, addressing accuracy and efficiency limitations in existing methods. It employs a dual-branch network to separately identify transient and periodic noise, combined with an improved spectral subtraction for precise audio separation. Model compression via pruning and knowledge distillation ensures real-time capability. Experimental results on AudioSet show recognition accuracies of 93.5% (transient) and 94.1% (periodic), with average denoising gains of 15.1 dB, 15.0 dB, and 16.3 dB for transient, periodic, and mixed noise, respectively. When processing 100 minutes of lab-recorded audio, latency remains under 21.0 ms, outperforming three benchmark models in speed and stability. The model demonstrates robust noise reduction and real-time performance, suitable for applications like live music, low-latency communication, high-quality post-production, and restoration of noisy historical recordings. Inderscience Publishers - linking academia, business and industry through research

Title: Computer music denoising and enhancement using dual-branch communication with spectral subtraction

Authors: Xiumei Ren

Addresses: College of Art, Jiaozuo University, Jiaozuo, Henan 454000, China

Abstract: This paper proposes a deep learning-based model for computer music denoising, addressing accuracy and efficiency limitations in existing methods. It employs a dual-branch network to separately identify transient and periodic noise, combined with an improved spectral subtraction for precise audio separation. Model compression via pruning and knowledge distillation ensures real-time capability. Experimental results on AudioSet show recognition accuracies of 93.5% (transient) and 94.1% (periodic), with average denoising gains of 15.1 dB, 15.0 dB, and 16.3 dB for transient, periodic, and mixed noise, respectively. When processing 100 minutes of lab-recorded audio, latency remains under 21.0 ms, outperforming three benchmark models in speed and stability. The model demonstrates robust noise reduction and real-time performance, suitable for applications like live music, low-latency communication, high-quality post-production, and restoration of noisy historical recordings.

Keywords: dual-branch communication; spectral subtraction; computer music denoising; model pruning; knowledge distillation.

DOI: 10.1504/IJICT.2026.153722

International Journal of Information and Communication Technology, 2026 Vol.27 No.54, pp.1 - 23

Received: 31 Jul 2025
Accepted: 15 Dec 2025
Published online: 21 May 2026 *

Title: Computer music denoising and enhancement using dual-branch communication with spectral subtraction

Keep up-to-date