Effective Noise Suppression in Whale Bioacoustics using Transfer-Learned Stacked BiLSTM

Abstract This study proposes a transfer learning-based noise suppression method employing a stacked bidirectional long short-term memory (BiLSTM) network for isolating killer whale vocalizations in complex underwater acoustic environments. The proposed framework employs a two-stage strategy, first using simplified synthetic waveforms for pretraining and then fine-tuning on field data. Input features are extracted via the fractional Fourier transform (FrFT), enabling improved feature separability. As a result, the FrFT angle range of − 67.5° to 67.5° achieved the best denoising performance, with a peak signal-to-noise ratio (PSNR) of 69.81 dB and a root mean square error (RMSE) of 3.23e–4. Comparative analyses on both synthetic and field datasets confirm that the proposed method effectively suppresses ambient and background noise while minimizing the distortion of essential bioacoustics signals of whales. This transfer learning approach demonstrated superior performance over conventional filtering techniques, which often degrade bioacoustic signals. These findings highlight the robustness of the method and its significant potential for enhancing marine mammal tracking and acoustic monitoring performance in marine science.