Data-driven Arabic phoneme recognition using varying number of HMM states

Continuous Arabic Speech Recognition, appears in many real life applications. Its speed, accuracy and improvement are highly dependent on the accuracy of the language phonemes set. The main goal of this research is to recognize and transcribe the Arabic phonemes based on a data-driven approach. We built a phoneme recognizer based on a data driven approach using HTK tool. Different numbers of Gaussian mixtures with different numbers of HMM states were used in modeling the Arabic phonemes in order to reach the best configuration. The corpus used consists of about 4000 files, representing 5 recorded hours of modern standard Arabic of TV-News. The maximum phoneme recognition accuracy reached was 56.79%. This result is very encouraging and shows the viability of our approach as compared to using a fixed number of HMM states.