Published on February 2017 | Artificial Intelligence
Background and Objective: Automatic speech recognition transcribes acoustic signal into strings of words for a given language. Speech recognition applications in native languages will enable information accessible for illiterate and disabled user in the society. In this research, the focus was on improving automatic speech recognition of the Tamil language. The development of a large-vocabulary continuous speech recognition system for Tamil language, which requires an acoustic model to be trained on a large vocabulary corpus. Methodology: To address the challenges, a modeling efficient sub-word units were recommended and designed a consonant-vowel six-segment (CVS-6) algorithm for syllabification of a Tamil text corpus and experimentally investigated its speech recognition accuracy. A specific database was constructed using 120 sentences of semi-continuous speech, comprising 561 words and 436 unique syllables. Results: The syllable-based model achieved a mean recognition rate of 81.41% (standard deviation, 6.94%) compared with the 69.87% (standard deviation, 4.11%) achieved by a phoneme-based model. The word error count for complex words was 25% by the syllable-based model compared with 54.96% by the phoneme-based model, a reduction of 30%. Conclusion: Syllable-based model using consonant-vowel six-segment algorithm is good choice and can be used to sub-word modeling of large vocabulary continue speech in Tamil language.