Home-environment adaptation of phoneme factorial hidden Markov models

Abstract

We focus on the problem of speech recognition in the presence of nonstationary sudden noise, which is very likely to happen in home environments. To handle this problem, a model compensation method based on a factorial hidden Markov model (FHMM) has been recently introduced. In this architecture, speech and noise processes are modeled in parallel by a phoneme FHMM that is built by combining a clean-speech phoneme hidden Markov model (HMM) and a sudden noise HMM. Here, to increase the robustness of this method further, we apply supervised and unsupervised home-environment adaptation of phoneme FHMMs. A database recorded by a personal robot PaPeRo in home environments was used for the evaluation of the proposed method under noisy conditions. The phoneme home-dependent FHMM achieved better recognition accuracy than the clean-speech home-independent HMM, reducing the overall relative error by 16.2% and 12.3% on average for supervised and unsupervised adaptation, respectively.