Data-Driven Speech Enhancement

Abstract:
Our new approach to speech enhancement employs a trained weighting rule for noisy speech spectral amplitudes. Training is performed with noise types of relevance for a specific application requiring separate clean speech and background noise training data. Training result is an individual weighting rule indexed by the a posteriori SNR and the a priori SNR for each frequency bin. The approach outperforms known speech enhancement techniques that perform explicit modeling of the speech and noise spectrum pdf. We prove the potential of our algorithm with simulations reporting noise attenuation and speech distortion.