Reduction of musical noise in spectral subtraction method using subframe phase randomisation

A subframe phase randomisation method is proposed and applied to the enhanced speech with spectral subtraction method to reduce musical noise in the nonvoicing region. The musical noise in the spectral subtraction method is largely due to narrowband tonal components appearing somewhat periodically in the spectrogram of unvoiced and silence regions. Thus, each synthesis frame in a nonvoicing region is divided into several subframes to broaden the narrowband spectrum, and then phases of silence and unvoiced regions are randomised to disrupt the tonal spectrum structure