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

Ocean noise negatively influences the recording of odontocete echolocation clicks. In this study, a hybrid model based on the convolutional neural network (CNN) and long short-term memory (LSTM) network-called a hybrid CNN-LSTM model-was proposed to denoise echolocation clicks. To learn the model parameters, the echolocation clicks were partially corrupted by adding ocean noise, and the model was trained to recover the original echolocation clicks. It can be difficult to collect large numbers of echolocation clicks free of ambient sea noise for training networks. Data augmentation and transfer learning were employed to address this problem. Based on Gabor functions, simulated echolocation clicks were generated to pre-train the network models, and the parameters of the networks were then fine-tuned using odontocete echolocation clicks. Finally, the performance of the proposed model was evaluated using synthetic data. The experimental results demonstrated the effectiveness of the proposed model for denoising two typical echolocation clicks-namely, narrowband high-frequency and broadband echolocation clicks. The denoising performance of hybrid models with the different number of convolution and LSTM layers was evaluated. Consequently, hybrid models with one convolutional layer and multiple LSTM layers are recommended, which can be adopted for denoising both types of echolocation clicks.

Transfer learning for denoising the echolocation clicks of finless porpoise (Neophocaena phocaenoides sunameri) using deep convolutional autoencoders.

Ocean noise has a negative impact on the acoustic recordings of odontocetes' echolocation clicks. In this study, deep convolutional autoencoders (DCAEs) are presented to denoise the echolocation clicks of the finless porpoise (Neophocaena phocaenoides sunameri). A DCAE consists of an encoder network and a decoder network. The encoder network is composed of convolutional layers and fully connected layers, whereas the decoder network consists of fully connected layers and transposed convolutional layers. The training scheme of the denoising autoencoder was applied to learn the DCAE parameters. In addition, transfer learning was employed to address the difficulty in collecting a large number of echolocation clicks that are free of ambient sea noise. Gabor functions were used to generate simulated clicks to pretrain the DCAEs; subsequently, the parameters of the DCAEs were fine-tuned using the echolocation clicks of the finless porpoise. The experimental results showed that a DCAE pretrained with simulated clicks achieved better denoising results than a DCAE trained only with echolocation clicks. Moreover, deep fully convolutional autoencoders, which are special DCAEs that do not contain fully connected layers, generally achieved better performance than the DCAEs that contain fully connected layers.