Hidden Markov Model Based Automatic Speech Recognition Using Mel Frequency Cepstral Coefficients In Nepalese

Nepalese (also called Nepali) is a language of some importance in the northern part of South Asia and is spoken mainly in Nepal, Bhutan and India. The impetus behind this undertaking to implement automatic speech recognition in Nepalese has been the fact that little research has been done in this area compared to the plethora of materials available for other languages like English. Hidden Markov models will be used with MFCC (Mel Frequency Cepstral Coefficients) analysis in the project. HMM, though applicable in many other pattern recognizers as well, has gained a prominent niche in ASR. The system, designed using HTK [1 HTKBook], starts with a preprocessing stage, which converts a speech waveform into feature vectors. The second stage is training the recognizer. Lastly, it will be used to decode new speech data. The building-block components of the system are phoneme-level statistical models. Word-level acoustic models will be formed by concatenating phone-level models according to a pronunciation dictionary. These word models will then be combined with a language model, which constrains the utterances to valid word sequences.