Is Phoneme Level Better than Word Level for HMM Models in Limited Vocabulary ASR Systems?

In this paper Arabic alphadigits were investigated from the speech recognition problem point of view. Limited vocabulary Arabic Automatic Speech Recognition Systems (ASRs) were designed, implemented, and tested by using isolated word utterances which consists of Arabic alphabets and/or digits. These systems were implemented separately by using phoneme level and word level based HMM models in distinct systems. The systems were isolated word speech recognizers with a database with size of 16,200 tokens in total that was created depending on 67 Arabic native speakers. Experiment were designed and investigated with different HMM models and different levels of viewing the sound units. Phoneme level and word level models were compared along with Arabic digits alone, alphabets alone, and, then, with both digits and alphabet. In all cases the whole system accuracies and individual utterance accuracies were generated and compared. This research showed that the phoneme level HMM models are superior for limited vocabulary ASR.