Isolated Spoken Marathi Words Recognition Using HMM

Automatic speech recognition is the computer driven conversion of spoken language into its corresponding text format. Hidden Markov Models (HMMs) are widely used statistical models in speech recognition systems. This paper presents a speaker independent word recognition system for Marathi language. Hidden Markov Model Toolkit (HTK) is used to implement the system. Mel Frequency Cepstral Coefficients (MFCCs) of phonetically rich 20 Marathi words collected from ten native speakers are used to train HMMs. Viterbi algorithm is used to recognize test word utterances. The performance of phoneme based HMMs and their recognition accuracy are discussed.