Lightweight neural networks for speech enhancement under drone noise: a comprehensive evaluation

Speech enhancement in the presence of drone noise presents a critical challenge due to the non-stationary nature of interference from motors, propellers, and wind. This paper conducts a comprehensive evaluation of lightweight neural networks for speech enhancement under drone noise, focusing on their ability to suppress noise while preserving speech quality. Seven lightweight models—Uformer, DCUNet, DCCRN, DPCRN, DPT-FSNet, FSPEN, and GTCRN—are assessed across various objective metrics, including PESQ, STOI, Si-SDR, and NISQA. The evaluation utilizes a dataset comprising clean speech from the TIMIT corpus combined with drone noise sourced from open datasets and custom recordings. A hybrid loss function, balancing magnitude loss and real-imaginary components, is applied to improve the models' performance in frequency domain reconstruction. Results show that DPT-FSNet achieves the best overall performance in low-SNR scenarios, demonstrating superior noise suppression with minimal distortion of intelligibility. However, the model struggles to maintain high-frequency speech details, highlighting a key area for improvement. These findings provide a detailed comparison of current lightweight architectures, offering valuable insights into their strengths and limitations when applied to real-world drone noise environments, and guiding future development in low-complexity models for noise-robust speech enhancement.