Acoustic Analysis of Indian Childrens Speech for Automatic Recognition of Mispronunciations and Speech Delay

Abstract

To overcome the problem of lack of speech samples of the Indian children, speech samples have been collected of children aged 5-15 years speaking English with accents of their mother tongue language. Furthermore, this work focuses on word level assessment in recorded children s speech and improving the communication in dyadic interactions. Word level assessment include word mispronunciations like, taking care of incorrect pronunciations or partially pronouncing word or combining words together. The advantage of the present work over other existing systems is primarily, that this model is trained on children speaking Indian English, which helps the system to adapt to their mispronunciations using Finite State Automata (FSA) in the Language model(LM). The LM helps to orient the utterances with the known transcripts and identify the word level pronunciation errors, as well as correct it. Also, during the annotation of the dataset, various possible disfluencies of the children are noted and the non-native phones and mispronunciations are included in the customised pronunciation dictionary which helps to handle variability in target Indian children. newline newline newline