Signal Representations For Audio Texture Synthesis Using Generative Adversarial Networks

 

 

This is a companion webpage for the paper - Signal Representations For Audio Texture Synthesis Using Generative Adversarial Networks

 

 

Our code is available on Github here

 

 

Simulating estimation error of IF

 

 

IF method is robust to estimation errors in temporally stable sounds with clearly separated frequency components, but is sensitive to estimation errors in short duration noisy wideband sounds with closely spaced frequency components.

 

 

 

 

 

 

Analysing GAN estimated outputs

 

 

We present a visual analysis of the GAN-estimated spectrograms using IF and PGHI methods.
Note: In IF method, the phase and phase gradient in frequency direction are estimated from the GAN estimated IF spectrogram. In PGHI method, the phase, phase gradient in frequency as well as time direction are all estimated from the GAN estimated log magnitude spectrogram

 

 

 

 

 

 

 

 

 

 

Examples of the generated audio used for the Listening Tests