Generative adversarial networks in sound synthesis: analysis of sound modeling capabilities using GANs.

This paper explores the capabilities of Generative Adversarial Networks (GANs) for audio synthesis. Particularly, it focus on WaveGAN, which works directly with raw waveforms, and DCGAN trained on spectrogram representations. These models were trained using internal combustion engine sounds collected from the Google AudioSet corpus. The study aims to assess how data representation and architectural choices affect not only quality but also training dynamics in synthesis. It exposed different behavioral characteristics of both approaches thus underlining the fact that signal representation largely defines the generative process and perceptual properties of output. Modelling on spectrograms, even though phase information is naturally lost, proved more robust and stable than waveform generation directly from waveforms. The results have further affirmed that a fair share of decisive synthesis quality with GAN-based audio synthesis resides in factors including data complexity, the domain of representation, and hyperparameter configuration.