A Generative Model for Cetacean Whistles Using Variational Autoencoder and Mixture of Gaussians

This paper proposes a tractable and compact generative model for cetacean whistle signals based on Variational Autoencoder (VAE) and mixture of Gaussians in underwater biomimetic communication. Cetacean whistles are nonlinear frequency-modulated signals with diverse time-frequency structures and can be interpreted as probability density functions. To learn the marginal probability distribution of whistle data, this study incorporates a VAE and a mixture of Gaussians. The VAE learns latent space representations from whistle spectrogram images, and the prior distribution of the latent variables is modeled using multiple Gaussian components. Since the latent space of whistle contour is modeled as an explicit density function of multimodal structure, it enables unsupervised clustering and conditional whistle generation by analyzing component-wise features. The statistical distance between the synthesized whistles and the training data was evaluated using the Fréchet inception distance, Recall, Precision, and F1-score yielding a score of 2.6767, 87.21%, 91.99%, and 89.54%, respectively, when the number of Gaussian components was set to 10,000.