Wind Noise Reduction Based on the Double Masking and Permutation-Invariant Training

Wind noise is a pervasive and non-stationary form of interference in outdoor audio recordings, posing a significant challenge for speech enhancement systems. To address this problem, this paper proposes a speech separation-based wind noise reduction framework termed Dual-mask permutation-invariant training (DMPIT). Building upon the dual-masking concept, the key contribution of DMPIT lies in embedding the dual-mask structure within a permutation-invariant training (PIT) framework and reformulating the loss function to better align with speech-oriented noise reduction objectives. Specifically, two supervised masks are jointly optimized: a speech mask that directly estimates the target speech from the mixture and a noise mask that isolates the wind noise component. Assuming that the mixture consists solely of speech and wind noise, the training process computes the loss using both estimated components and the corresponding clean speech. Since wind noise is not a signal of interest, the estimated noise is subtracted from the mixture to obtain a residual speech signal, which is then used to refine the direct speech estimate. The final enhanced speech output is produced by fusing the direct and residual speech estimates through a weighted combination. The experimental results demonstrate that DMPIT consistently outperforms conventional single-mask and single-channel wind noise reduction methods in terms of speech quality and noise suppression.