Using AR HMM state-dependent filtering for speech enhancement

In this paper we address the problem of enhancing speech which has been degraded by additive noise. As proposed by Ephraim et al., autoregressive hidden Markov models (AR-HMM) for the clean speech and an autoregressive Gaussian for the noise are used. The filter applied to a given frame of noisy speech is estimated using the noise model and the autoregressive Gaussian having the highest a posteriori probability given the decoded state sequence. The success of this technique is highly dependent on accurate estimation of the best state sequence. A new strategy combining the use of cepstral-based HMMs, autoregressive HMMs, and a model combination technique, is proposed. The intelligibility of the enhanced speech is indirectly assessed via speech recognition, by comparing performance on noisy speech with compensated models to performance on the enhanced speech with clean-speech models. The results on enhanced speech are as good as our best results obtained with noise compensated models.