An Implementation of Text Dependent Speaker Independent Isolated Word Speech Recognition Using HMM

Speech is the powerful tool of information exchange. There are different aspects related to speech like speech recognition, speech verification, speech synthesis, speaker recognition, speaker identification etc. Speech Recognition is the process of determining which speech is present based on individual’s utterance. This paper gives an implementation& performance evaluation of Text Dependent Speaker Independent Isolated Word Speech Recognition System is using HMM. Here MFCC is used for speech parameterization. These feature parameters are used for HMM training. In HMM modeling forward backward algorithm with EM principle is used for parameter estimation & optimization. Initially K means algorithm is used for dividing feature dataset into smaller parts and then means are calculated.GMM is used to model the distribution of speech features for each state of HMM. Finally the computed HMM parameters for words are stored in respective HMM models as a reference database. To recognize the spoken word the likelihood of generation of test speech observations and the most likely path sequence through each stored HMM model models is calculated. Finally the one with maximum likelihood path is selected as recognized word. The system is implemented in MATLAB 7.9.The system is trained with own created database consisting 60 speech samples of selected words & is tested for speaker independent mode. For selected three words recognition accuracy of 92%, 92% & 88% respectively is obtained in noisy environment.