Bangla pronunciation error detection system

In this paper a pronunciation error detection system was modeled and also tested for large vocabulary, speaker independent and continuous speech recognizer for Bengali language. The recognizer was developed using Hidden Markov Model (HMM); and the Hidden Markov Modeling Toolkit was used to implement it. In the process, a corpus database comprised of 3000 utterances that were used for training and 100 plus sentences for development and evaluation. The data was preprocessed in line with the requirements of the HTK toolkit. In order to support the acoustic models, a bigram language model was constructed. In addition, pronunciation dictionary was prepared and used as an input. Standard experiment for in depth performance analysis of the detector was designed and all the results are analyzed to get the proper insight of the error pattern, both at sentence and word level. The effect of Conjugant words, the effect of first letter of the word, the gender effect on error pattern along with the effect of colloquial dialect or local accent on error pattern is observed. The findings are quite promising and may open new possibilities to design an efficient pronunciation error correcting system for aiding the non-native Bengali speaking people.