Using Linguistic Data in a Continuous Speech Recognition System for Persian

In this paper, a continuous speech recognition system for the Persian language is introduced and the roles of acoustic and language models are examined. Context-independent and context-dependent acoustic models are used in the system and the results of their employment are presented. Moreover, word-based, POS-based and class-based triphone language models are extracted using Persian text corpus and incorporated in the speech recognition system. In addition, a grammatical language model based on GPSG is implemented in the system and is used in combination with the statistical language model. Experimental results demonstrated hat as expected, context-dependent phonetic models show the best performances. Also, the word-based triphone language model showed superiority over other statistical language models. Moreover, the combination of grammatical language models with statistical ones proved to lead to better recognition results. The system introduced in this paper is the first Persian speech recognition system capable of practical usage and is based on numerous research works performed for its design and implementation.