Combining NDHMM and phonetic feature detection for speech recognition

Abstract

Non-negative HMM (N-HMM) [1] is a model that is well suited for modeling a mixture of e.g. audio signals, but does not have the ability to generalize to model unseen data. Non-negative durational HMM (NdHMM) has recently been proposed [2] as a modification to N-HMM that can allow for generalization, and thus make the approach suitable for automatic speech recognition. A detector-based approach to speech recognition has been studied by several researchers as an alternative to the traditional HMM approach. A bank of phonetic feature detectors will produce phonetic feature posteriors, which fit well with the non-negativity constraint of NdHMM. We review the NdHMM approach proposed in [2] and propose to extend this approach by combining NdHMM with a phonetic feature detection front-end in a tandem-like system. Experimental results of the proposed approach are presented.