Noisy speech enhancement with sparsity regularization

In this paper, a novel unsupervised speech enhancement algorithm is proposed assuming that both speech spectrogram and its temporal gradient are sparse. This assumption is reliable due to quasi-harmonic nature of speech signals. In the proposed method, speech enhancement is performed by minimizing an appropriate objective function composed of a data fidelity term and sparsity imposing regularization terms. Alternating direction method of multipliers (ADMM) is adapted to solve the proposed model, and an efficient iterative algorithm is developed for speech enhancement. Extensive experiments demonstrate that the proposed method outperforms other competing methods in terms of different performance evaluation metrics.