Despite artificial neural networks (ANNs) making rapid progress in the field of noise removal for audio signals, their computational complexity and unpredictable behavior on unseen noise types constitute an issue for many applications. Noise suppression systems often need to be adopted in low-resource communications systems that cannot meet the requirements of most deep learning models. These systems also require real-time low-delay processing, and are adopted in a wide variety of noise situations. To overcome these limitations, we propose an innovative hybrid noise suppressor (HNS) for speech signals which combines the robustness of a traditional stationary noise suppressor (SNS) with the generalization capabilities of an ANN. A low-complexity ANN is used to enhance the performance of the SNS by removing non-stationary noise and perceptually unpleasant artifacts. Our evaluation shows that the proposed HNS is able to perform effective real-time denoising on unseen noise types, while retaining a lower complexity than the vast majority of state of the art deep learning techniques.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=22538
(1866KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
