Audio Image Generation for Denoising

Jialu Li, Youshan Zhang

Keywords: Audio Denoise; Diffusion models

Abstract: Time-frequency domain analysis has emerged as an effective method to remove noise in audio signals. However, the image generation quality of the frequency domain is not yet well explored. In this paper, we turn the audio denoising task into an image generation problem. We present an audio image generation model for audio denoising named AIGD and use it to estimate the posterior distribution of clean complex images conditioned on noisy complex images. Given any noisy audio signals, our AIGD model could directly generate denoised complex images and output clean audio signals. We further optimize complex L2 and complex absolute structure similarity losses to improve the quality of generated images. An SDR loss is proposed to reconstruct better-denoised audios. Extensive experimental results demonstrate that by generating high-quality frequency domain images, our AIGD model achieves state-of-the-art performance audio denoising.

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5673