A soft-suppression approach for speech enhancement under a deterministic noise model

Although many single-channel based speech enhancement methods rely on stochastic models to estimate noise, certain speech-like noise clearly shows a more deterministic character. In the paper, we study a soft-suppression approach for enhancing speech corrupted by speech-like noise under a deterministic noise model. High performance is achieved by following two stages: In the first stage, the frequency span of speech-like noise is pre-estimated under a deterministic model. The second stage employs a novel externally-linear-internally-nonlinear (ELIN) system to realize soft-suppression and enhance noisy speech. Thanks to the soft-suppression method, we are able to improve the quality of speech without losing detail information which is significant in speech perception. The simulation results show that the proposed approach can operate effectively even under low SNR conditions.