Adversarial Generation ofTime-Frequency Features

with application in audio synthesis

Speaker: Andrés MarafiotiCo-Authors: Nathanaël Perraudin, Nicki Holighaus, Piotr Majdak

Acoustics Research Institute, ViennaAustrian Academy of Sciences

International Conference on Machine LearningLong Beach, California, June 11th, 2019

Time to time-frequency

Marafioti (ARI) Adversarial Generation of TF Features ARI 2 / 6

Time-frequency to time

Marafioti (ARI) Adversarial Generation of TF Features ARI 3 / 6

Is it consistent?

Marafioti (ARI) Adversarial Generation of TF Features ARI 4 / 6

Applied to GANs

Marafioti (ARI) Adversarial Generation of TF Features ARI 5 / 6

Evaluation

We trained on a dataset of spoken English digits [0-9].We evaluated our results with perceptual tests.Audio examples and implementations are available attifgan.github.io

WaveGAN digits TiFGAN-M digits

vs TiFGAN vs WaveGANReal 86% 94%TiFGAN – 75%WaveGAN 25% –

Thank you for your attention!Supported by the Austrian Science Fund

(FWF; MERLIN, I 3067-N30).

Marafioti (ARI) Adversarial Generation of TF Features ARI 6 / 6

tifgan.github.io

Problem studied

fd@rm@0: fd@rm@1: