Spatiotemporal Averagingmethod for Enhancement of Reverberant Speech

We present a reverberant speech enhancement algorithm which, operating on the linear prediction residual of spatially averaged multi- microphone observations, utilizes temporal averaging of neighbouring larynx cycles. The enhanced larynx cycles are used to design an equalization filter, which is applied in order to dereverberate both voiced and unvoiced speech. The DYPSA algorithm is employed and evaluated for larynx cycle segmentation of reverberant speech. Simulation results show reverberation reduction with up to 5 dB in terms of segmental signal-to-reverberant ratio and 0.34 in terms of normalized Bark spectral distortion score.