Reconstruction modeling and validation of brown croaker ( Miichthys miiuy ) vocalizations using wavelet-based inversion and deep learning

Biological vocalizations of fish serve as essential acoustic cues for underwater monitoring and ecological studies. However, limited high-quality recordings, especially for region-specific species like the brown croaker (Miichthys miiuy), hinder data-driven modeling in bioacoustics. This study introduces a wavelet-based synthesis framework using fk14 wavelets and Particle Swarm Optimization (PSO) to reconstruct brown croaker calls. A total of 522 single-pulse and 1157 pulse train signals were analyzed to extract key acoustic parameters such as sound pressure level (SPL0−pk�) and inter-pulse interval (IPI). Sensitivity analysis using a normalized Bartlett processor identified wavelet delay and scale as the most influential parameters, guiding the generation of biologically valid signals with over 98% waveform similarity. The synthesized calls closely matched measured signals in both time and frequency domains. A ResNet-18 neural network trained on measured signals successfully distinguished between measured and modeled calls, demonstrating high classification performance and strong similarity in latent feature space. These findings confirm the structural fidelity of the synthetic signals and demonstrate the potential of this method for data augmentation, passive acoustic monitoring, and the development of bio-inspired underwater communication systems. (Work supported by KRIT-CT-22-056.)