Denoising odontocete echolocation clicks using a hybrid model with convolutional neural network and long short-term memory network.

This article focuses on the development and evaluation of a hybrid convolutional neural network–long short-term memory (CNN-LSTM) model designed to denoise odontocete echolocation clicks affected by ocean noise. The model was pre-trained using simulated echolocation clicks generated from Gabor functions and fine-tuned with real echolocation data from finless porpoises (narrowband high-frequency clicks) and Irrawaddy dolphins (broadband clicks). Experimental results demonstrated that bidirectional LSTM (BLSTM) layers outperform unidirectional LSTM (ULSTM) layers, particularly for longer-duration clicks, and that increasing the number of LSTM layers enhances denoising performance more than adding convolutional layers. The study recommends a hybrid CNN-LSTM model with one convolutional layer and multiple LSTM layers as an effective approach for denoising both types of odontocete echolocation clicks recorded via passive acoustic monitoring.