Alphabet model-based short vocabulary speech recognition for the assessment of profoundly deaf and hard of hearing speeches

Abstract

Speech quality will be degraded if any one of the cavities such as vocal, nasal, mouth or oral is imperfect. Even though the cavities are in good condition the children who have problems in the ear cannot reproduce sounds since they cannot hear. Their speech characteristic in terms of recognition accuracy is analysed for nine children in the age group of 10-14 years in their native classical Tamil language. Short vocabulary is considered for this purpose and based on this continuous speech monophone-based and senone-based speech recognition systems are developed. To capture the individual performance, speaker independent models are evaluated using nine-fold cross validation. It is observed that senone-based system performs well for them and they can be categorised as profoundly deaf and hard of hearing depending on their recognition accuracy. But some of them outperform well even though they are profoundly deaf since they have undergone speech therapy earlier. It is further observed that if phoneme model is replaced by simple alphabet model it reduces the system complexity and increases the recognition accuracy in average by 9.57%. Compared to clinical assessment, the present status of hearing impairment is well analysed by using the proposed speech recognition system.