A Minimum Mean Square Error Approach for Speech Enhancement

The statistical model based approach for enhancing noisy speech proposed in [1]-[3], comprises the estimation of the unknown probability distributions (PD's) of the speech signal and the noise process from long training sequences, and the application of the estimated PD's for constructing the desired estimators of the clean speech signal. The estimation of the PD's of the speech signal and the noise process is performed by first attributing hidden Markov models (HMM's) to the two processes, and then estimating the parameter sets of the models from the given training sequences.