Allophone-based acoustic modeling for Persian phoneme recognition

Phoneme recognition is one of the fundamental phases of automatic speech recognition. Coarticulation which refers to the integration of sounds, is one of the important obstacles in phoneme recognition. In other words, each phone is influenced and changed by the characteristics of its neighbor phones, and coarticulation is responsible for most of these changes. The idea of modeling the effects of speech context, and using the context-dependent models in phoneme recognition is a method which used to compensate the negative effects of coarticulation. According to this method, if two similar phonemes in speech have different contexts, each of them constitute a separate model. In this research, a linguistic method called allophonic modeling has been used to model context effects in Persian phoneme recognition. For this purpose, in the first phase, the rules required for occurrence of various allophones of each phoneme, are extracted from Persian linguistic resources. So each phoneme is considered as a class, consisting of its various context-dependent forms named allophones. The necessary prerequisites for modeling and identifying allophones, is an allophonic corpus. Since there was no such corpus in Persian language, SMALL FARSDAT corpus has been used. This corpus is segmented and labelled manually for each sentence, word and phoneme. So the phonological and lingual context required for the realization of allophones, is implemented in this corpus. For example, the syllabification has been performed on the corpus and then, for each phoneme, its position (first, middle and end) in the word and syllable is specified using different numeric tags. In the next step, allophonic labeling has been performed by searching for each of the allophonic contexts in the corpus. These allophonic corpus is used to model and recognize the allophones of input speech. Finally, each allophone is assigned to a proper phonemic class so phoneme recognition has been done using allophones. The experimental results show a high accuracy of the proposed method in phenome recognition, indicating a significant improvement comparing with other state-of-the-art methods.