The development of deep convolutional generative adversarial network to synthesize odontocetes' clicks

Odontocetes are capable of dynamically changing their echolocation clicks to efficiently detect targets, and learning their clicking strategy can facilitate the design of man-made detecting signals. In this study, we developed deep convolutional generative adversarial networks guided by an acoustic feature vector (AF-DCGANs) to synthesize narrowband clicks of the finless porpoise (Neophocaena phocaenoides sunameri) and broadband clicks of the bottlenose dolphins (Tursiops truncatus). The average short-time objective intelligibility (STOI), spectral correlation coefficient (Spe-CORR), waveform correlation coefficient (Wave-CORR), and dynamic time warping distance (DTW-Distance) of the synthetic clicks were 0.975, 0.968, 0.877, and 0.992, respectively. AF-DCGAN outperformed the minimum phase signal reconstruction (MPSR) method and variational quantized variational autoencoders (VQ-VAE) by 5.9% and 3.7% in STOI, 5.2% and 3.5% in Spe-CORR, and 5.8% and 2.8% in Wave-CORR, respectively. In addition, AF-DCGAN reduced DTW-Distances by 29.9% and 9.4% compared to MPSR and VQ-VAE, respectively. Results showed that AF-DCGAN was robust in synthesizing both narrowband and broadband clicks that can produce a substantial number of high-fidelity odontocetes' clicks with flexibility in modulating parameters. Employing AF-DCGAN to synthesize odontocete-like clicks could advance the development of a click database, offering promising applications in the research of biomimetic target detection and recognition.