An HMM system for recognizing articulation features for Arabic phones

In this paper, we introduce a Hidden Markov Model (HMM) recognition system for the articulation features of Arabic phones. The low-level features are described by Mel-Frequency Cepstral Coefficients (MFCCs). The created HMMs directly model certain articulation features (fricative and plosive). Classification is done on these features regardless of the phone itself. The model has been created successfully and tested on reference speech data. The error rate is very low for many phones and acceptable for most of them. Accordingly, the system output can be used as a confidence measure applied to other existing speech recognizers. Finally, the recognizer is speaker-independent and context-independent.