Musical residual noise is a major problem for a speech enhancement system. In this paper, we attempt to reduce the effect of musical residual noise by merging homogeneous spectral bins in an analysis block to mitigate the musical effect. Initially, we analyze the motion vector of each spectral bin. A motion difference which indicates a spectral bin being musical tone is then employed to adapt the merging threshold. In the case of a musical tone, the larger the value of the motion difference is, the more the number of the merged spectral bins is. It enables the spectral bins of a musical tone to vary smoothly over successive frames and among neighbour subbands. The effect of musical residual noise is accordingly mitigated. Experimental results show that the proposed post processing approach can efficiently reduce the effect of musical residual noise, while the speech quality can be well maintained.
IASTED Advances on Computer Science and Engineering (ACSE)