In this paper we investigate an incorporation of mask modeling within the missing-feature HMM-based ASR and also explore on feature representation within this framework. The mask model is estimated for each HMM state and mixture by using a separate Viterbi-style training procedure. We explore an employment of the frequency-filtered features and their combination with the logarithm filter-bank energies. Experimental evaluation is performed on the Consonant Challenge corpus. The obtained results show significant improvements by incorporation of the proposed methods over the standard MFT-based ASR.
Bibliographic reference. Jančovič, Peter / Kokuer, Munevver (2008): "On the mask modeling and feature representation in the missing-feature ASR: evaluation on the Consonant Challenge", In INTERSPEECH-2008, 1777-1780.