GAN-based Augmentation for Populating Speech Dataset with High Fidelity Synthesized Audio

In this paper, we present an audio augmentation method that generates synthetic audio using Generative Adversarial Networks (GANs). We propose a training strategy that first uses Harmonic Percussive Source Separation (HPSS) to extract spectral features and then improves the fidelity of the synthesized audio by applying progressively-growing GANs. Our method is demonstrated on a public speech dataset released by Google TensorFlow. When employing our method, the performance evaluated by Frèchet Inception Distance (FID) showed 12.514, but 14.012 as for the existing image-generating GANs (lower FID indicates better fidelity).