This paper proposes a novel delayless transient noise reduction method that is based on speech properties and spectral coherence. The proposed method has three stages. First, the transient noise components are detected in each subband by using energy-normalized variance. Second, we apply the harmonic property of the voiced speech and the continuity of the speech signal to reduce speech distortion in voiced speech segments. Third, we define a new spectral coherence to distinguish the unvoiced speech from the transient noise to avoid suppressing the unvoiced speech. Compared with those existing methods, the proposed method is computationally efficient and casual. Experimental results show that the proposed algorithm can effectively suppress transient noise up to 30 dB without introducing audible speech distortion.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.