Particle Filter Based Soft-mask Estimation for Missing Feature Reconstruction

In this work, we show how particle filter (PF) based speech feature enhancement can profitably be combined with soft-decision missing feature reconstruction. The combined approach is motivated by the fact that standard minimum mean square error noise compensation techniques fail to give accurate estimates of the clean speech spectrum if the noise spectral power significantly exceeds that of speech in a particular spectral region. Experiments show that the proposed algorithm can reduce the word error rate by up to 26.1% relative, compared to 17.0% for speech feature enhancement based solely on particle filters.