Automatic Recognition of Hearing-Impaired Children's Kannada Speech with Limited Vocabulary

Hearing is essential in our daily lives, helping us to understand language through sound and contributing to proper speech development. Hearing loss affects many people worldwide, and children account for a large part which can hinder their communication and development skills. Hearing impairment can influence the progression of children's participation in activities suitable for their age, as well as their functional speech communication abilities and language proficiency. Kannada database developed in collaboration with the All India Institute of Speech and Hearing (AIISH), Mysore, is used for hearing-impaired (HI) children. The main aim of this work is to build an automatic speech recognizer (ASR) system with a limited vocabulary dataset using Raspberry Pi. Data augmentation is used since the dataset is limited in number of words and number of children. Shorttime Fourier transform (STFT) spectrograms are extracted by varying frame sizes by incorporating augmentation to build an ASR model. A convolutional neural network (CNN) classifier is used to build an ASR model. The speech recognition accuracies for HI speech are compared with and without data augmentation. Improved recognition accuracy is obtained with data augmentation. Finally, real-time implementation for ASR is also carried out and it is observed that there is a scope for improvement of the developed ASR system.