Single-Channel Impact-Noise Suppression with no Auxiliary Information for its Detection

This paper proposes impact-noise suppression algorithm with no auxiliary information for its detection. Impact noise is successfully detected based on the shape and the change of the power spectrum. The impact-noise components are either replaced by an ambient-noise estimate or subtracted using an impact-noise estimate depending on the frequency and voice activity. Subjective evaluation results demonstrate that the proposed impact-noise suppression algorithm provides as much as 1.1 higher score in the 5-grade MOS than a nonimpact noise suppression algorithm and no suppression with statistically significant difference.