Arabic speech recognition using Hidden Markov Model Toolkit(HTK)

In this paper we discuss the development and implementation of an Arabic automatic speech recognition engine. The engine can recognize both continuous speech and isolated words. The system was developed using the Hidden Markov Model Toolkit. First, an Arabic dictionary was built by composing the words to its phones. Next, Mel Frequency Cepstral Coefficients (MFCC) of the speech samples are derived to extract the speech feature vectors. Then, the training of the engine based on triphones is developed to estimate the parameters for a Hidden Markov Model. To test the engine, the database consisting of speech utterance from thirteen Arabian native speakers is used which is divided into ten speaker-dependent and three speaker-independent samples. The experimental results showed that the overall system performance was 90.62%, 98.01 % and 97.99% for sentence correction, word correction and word accuracy respectively.