Conditioning Trick for Training Stable GANs

TL;DR

This paper introduces a novel conditioning trick for GAN training that improves stability and quality of generated audio spectrograms by aligning generator outputs with spectral domain properties of real samples, enhancing performance on environmental sound datasets.

Contribution

The paper proposes a new conditioning trick based on spectral domain analysis to stabilize GAN training and improve audio synthesis quality, integrating residual networks into BigGAN architecture.

Findings

Outperforms baseline GANs on multiple metrics

Achieves higher inception scores and lower FID

Produces high-quality, phase-preserving audio spectrograms

Abstract

In this paper we propose a conditioning trick, called difference departure from normality, applied on the generator network in response to instability issues during GAN training. We force the generator to get closer to the departure from normality function of real samples computed in the spectral domain of Schur decomposition. This binding makes the generator amenable to truncation and does not limit exploring all the possible modes. We slightly modify the BigGAN architecture incorporating residual network for synthesizing 2D representations of audio signals which enables reconstructing high quality sounds with some preserved phase information. Additionally, the proposed conditional training scenario makes a trade-off between fidelity and variety for the generated spectrograms. The experimental results on UrbanSound8k and ESC-50 environmental sound datasets and the Mozilla common voice dataset have shown that the proposed GAN configuration with the conditioning trick remarkably outperforms baseline architectures, according to three objective metrics: inception score, Frechet inception distance, and signal-to-noise ratio.