A Knowledge Distillation Framework for Multi-Signal-to-Noise Ratio Speech Enhancement in Construction Environments

Construction noise severely disrupts mission-critical communications on construction sites, significantly hindering operational efficiency and safety. However, existing speech enhancement methods fail to address two domain-specific challenges: processing speech under extreme signal-to-noise ratio (SNR) volatility, and overcoming mission-critical dataset scarcity constraints due to hazardous on-site conditions. To tackle these issues, this paper proposes an SNR-aware knowledge distillation framework tailored for construction environments. The framework trains specialized teacher models on SNR-segmented datasets, enabling a student model to learn multi-range denoising strategies through knowledge distillation. Additionally, a two-stage training strategy with domain-adaptive generalization is proposed to reduce dependence on diverse construction-domain samples, while a dynamic weight mechanism is designed to mitigate cross-domain shift. Experimental results demonstrate that the proposed method achieves average improvements of 0.01 in perceptual evaluation of speech quality (PESQ) and 0.23 in scale-invariant signal-to-noise ratio (SI-SNR) metrics compared to a non-distilled variant of the framework. This approach enables specialized acoustic applications for construction communication devices.