Effect of states and mixtures in HMM model and Connected word Recognition of Profoundly deaf and hard of hearing speech

摘要

It is a challenge for many years that how to fix the no. of states and no. of mixtures when HMM models are used for speech recognition. In this paper we have analysed that for hearing impaired speech that is partially intelligible to people who are speaking to them frequently and it is not understandable by the unfamiliar listeners. They suffer in many aspects like education and in public places to converse with the normal speakers. Since speech is unique most of the time normal speech itself could not be understand by others. If we develop the speech recognizer for their speech it will convert their unintelligible speech into intelligible speech. Speaker dependent connected digit recognition for this task using HTK tool kit is done and the average recognition accuracy obtained is 93%. Totally 10 speakers out of which 3 are hard of hearing and 7 are profoundly deaf are considered for this experiment. Then for isolated words, no. of mixtures are varied from 3 to 10 for each set of states such as 6, 7, 8, 9, 10 and recognition accuracy is verified for each case. When we varied beyond that there is no any significant change in recognition accuracy and so it is concluded that we can have mixture and state value as 10 for small vocabulary and the recognition performance for all types of feature is comparable to that of normal speech recognition. But irrespective of the state higher recognition is achieved at 8 or 9 or 10 mixer value for different type of feature and it can be concluded that, if we have the mixer value as 8 , 9 or 10 we can get reasonable results.