Self-Supervised Speech Denoising Using Only Noisy Audio Signals

In traditional speech denoising tasks, clean audio signals are often used as the training target, but absolute clean signals are collected from expensive recording equipment or studios with strict environment. To address this issue, we propose an end-to-end self-supervised speech denoising training scheme using only noisy audio signals named On-ly-Noisy Training (ONT), overcoming the limitation of clean speech collection without any extra training condi-tions. The proposed ONT constructs training pairs only from each single noisy audio, and it contains two modules: training audio pairs generated module and speech de-noising module. The first module adopts a random audio sub-sampler on each noisy audio to generate training pairs. The sub-sampled pairs are then fed into a novel com-plex-valued speech denoising module. Experimental results show that the proposed method not only eliminates the high dependence on clean targets of traditional audio denoising tasks, but also achieves on-par or better performance than other training strategies. Source code was released in https://github.com/liqingchunnnn/Only-Noisy-Training