Abstract

Leakage of water pipelines in urban underground utility tunnels poses a major threat to tunnel safety and operation; therefore, voiceprint recognition is adopted for real-time pipeline monitoring. Although utility tunnels are less affected by outdoor interference, multiple internal noises still degrade voiceprint recognition accuracy. To address this problem, this study proposes an enhanced StarGAN-based denoising method using a single network to handle multiple noise types. Unlike the original StarGAN-VC2 developed for voice conversion, the proposed model is specifically redesigned for leakage voiceprint denoising by integrating MFCC-based representation, a lightweight bottleneck, channel attention, residual feature preservation, and U-Net-style reconstruction. Experimental and engineering application results show that the denoised signals achieve improvements of 3–7 dB in SNR, 3–4 dB in PSNR, and 3–4 in SSR. The model also demonstrates strong generalization capability and plug-and-play applicability, enabling integration with conventional denoising and voiceprint recognition networks. These results indicate that the proposed method can effectively suppress diverse utility tunnel noises while preserving leakage-related voiceprint features.

IPC Classification

Keywords