Tree-Based HMM State Tying for Arabic Continuous Speech Recognition

One of the major challenges in building Hidden Markov Models (HMMs) for continuous speech recognition systems is the balance between the available training set and the recognition performance. For large vocabulary recognition systems, context dependent models are usually required to obtain higher recognition accuracy. This is crucial as most of the language contexts may not occur in the training set. This paper proposes an Arabic phonetic decision tree necessary to build tied state tri-phone HMMs. Experimental results based on the proposed decision tree show a promising recognition accuracy when compared with the traditional context independent models using the same training and testing sets. The maximum recognition accuracy achieved by the proposed approach was 92.8 % whereas it reached 61.5 % when tested using context independent HMMs.