Deep Speech Denoising with Minimal Dependence on Clean Speech Data

Most of the existing deep learning-based speech denoising methods rely heavily on clean speech data. According to the traditional view, a large number of noisy and clean speech samples are required for good speech denoising performance. However, the data collection is a technical barrier to this criteria, particularly in economically challenged areas and for languages with limited resources. Training deep denoising networks with only noisy speech samples is a viable option to avoid dependence on sample data size. In this study, the target and input of a DCU-Net were trained using only noisy speech samples. Experimental results demonstrate that, when compared to traditional speech denoising techniques, the proposed approach avoids not only the high dependence on clean targets but also the high dependence on large data sizes.