Comparative experiments to evaluate the use of syllables for large-vocabulary automatic speech recognition

This paper motivates the use of syllables to enhance the performance of automatic speech recognition (ASR) systems when dealing with large-vocabulary speech. Arabic and English are considered in our paper to test the proposed approach. The Arabic database consists of sentences selected from different Arabic broadcast news, whereas for English speech, TIMIT database had been used to test our approach. Comparative experiments have indicated that the use of syllables as acoustic units for the recognition of both languages leads to an improvement in the recognition performance of HMM-based ASR systems. The Hidden Markov Model Toolkit (HTK) was used throughout our experiments. A series of experiments on speaker-independent continuous-speech recognition have been carried out using both databases. Using such an approach, experiments show that for Arabic database, the recognition rate using syllables outperforms the recognition rate obtained using monophones and triphones by 15.75% and 2.64%, respectively. On the other hand, for TIMIT database, the recognition rate using syllables outperforms the recognition rate using monophones and triphones by 40.08% and 19.74%, respectively.