coherence weighting use magnitude squared coherence(MSC) supression gain:
limit gain function
coherence based wind noise estimation wind noise estimation and reduction separation. combine with other disturbance estimation. dual microphone signal model in FFT domain:
for uncorrelated noise signal,MSC power spectrum:
an estimation of the noise PSD:
short-term estimates of the PSD used recursive smoothing approach:
the smoothing process for the computation of the PSDs lead to an slow adaption. smoothing factor=0.5,an overestimation of the MSC and thus of the cross-PSD laed to an underestimation of the wind noise. smoothing factor=0.96,an underestimation of the MSC , lead to a too high wind noise estimate.
for a coherent signal the phase difference is only dependent on the DOA of this signal.
u_c should be chosen to a range in which both wind noise and speech are active,eg …,to 0…4000Hz.
Adaptive smoothing factor for improved coherence
the adaptive smoothing factor limited to the range 0.5——1. 3.