Speech recognition system for Tamil language using Cmusphinx

Abstract

Speech recognition is the process of converting an audio signal into sequences of words paving the way for enabling a rich human-computer interaction in many emerging applications. It is a very difficult and a complex task to recognize speech using a computer though several approaches have been made to develop an accurate speech recognition system. Most of the researches have been done for English, Chinese, mandarin and Arabic languages while a very few have been done for Tamil language. Therefore, this research intends to develop a continuous speech recognition system for Tamil language. Tamil language has 12 vowels and 18 consonants. This paper is focused on building a Tamil speech recognition system using Cmusphinx toolkit. The system was developed for several desktop applications with the aim that in the future, this system could be integrated with those applications and executes those applications through Tamil speech input rather than typing and Clickin. Audio recordings from several people were recorded in different environment which were later converted to wav file format. A text corpus, transcription file and fileids file for both training and testing database were prepared accordingly. Three models namely acoustic model, language model and lexicon model have been developed. Hidden Markov model was employed for building acoustic model. The performance of the system over various speaker subsets of different sex, age and dialect was examined. The word error rate was calculated in order to measure the performance of the system while the accuracy of the system was calculated using another formula. The results from both the measurements showed a satisfactory performance.