Robust speech enhancement based on NPHMM under unknown noise

In this paper, a new speech enhancement based on the nonlinear H-infinity filtering and neural predictive HMM (NPHMM) is presented. In H-infinity filtering, no a prior knowledge of the noise source statistics is required. Speech is modeled as the output of a neural predictive HMM combining MLP neural network and HMM. The proposed enhancement method consists of multiple nonlinear H. filters with parameter of the NPHMM. The switching between the nonlinear H. filters is governed by a finite state Markov chain according to the transition probabilities. An approximate improvement of 0.4-1.8dB in output SNR is achieved at various input SNRs compared with conventional Kalman method.