Harmonic Structure Mask for Speech Enhancement Using Sparsity Regularization

Harmonic structure, an important characteristic of speech signals, has been utilized in various speech processing applications, such as dereverberation, fundamental frequency (f
₀
) estimation, voice activity detection (VAD), phase reconstruction and source separation. This paper presents a harmonic structure mask for those speech enhancement applications based on sparsity regularization via convex optimization. Specifically speaking, we first derive a harmonic structure mask of the noisy speech using f
₀
and VAD estimations, then use this mask to protect harmonic components of speech during the sparsity regularization process. The proposed mask benefits from the additional harmonic information, leading to better protection of harmonic components. Numerical experiments show that the proposed mask can improve speech quality and intelligibility compared to the previous work.