Detection And Classification of Childhood Apraxia of Speech Using Deep Guided Convolution Neural Network

Abstract

Childhood Apraxia of speech is neurological motor speech disorder which is due to difficulty of brain in planning and programming the complex movement of the speech. Especially it can’t be categorized on basis of muscle weakness.  Thus, it becomes mandatory to design a speech recognition system towards detection of childhood apraxia of speech on recorded sounds of child and doctor conversation. However, manual speech processing technique becomes challenging to classify the childhood apraxia of speech due to its complex sampling rate. Adoption of machine learning architecture from artificial intelligence to speech recognition makes detection more feasible and accurate. Despite of several advantage of the machine learning and deep learning approaches, there exist some challenges on basis of model scalability to large vocabulary and speech variability due to accent and style. In order to mitigate those challenge, deep learning model has to be modelled. In this paper. a new deep guided convolution neural network is designed and implemented to classify Childhood Apraxia of speech. Initially preprocessing step is performed to eliminate the noise and transform signal into segmented frame. Next segmented frame is processed in fast Fourier transform to obtain the power spectrum. Obtained Power spectrum is projected to proposed model. Convolution layer of model use mel filter to MFCC features and it is organized in feature map. Extracted feature is employed to fully connected network to perform precise recognition and classification of the Childhood Apraxia of speech in order to enhance prognosis of the specified disease. Experimental analysis and performance analysis of the proposed model have been evaluated using speech dataset from Ultra Suite Repository in the Python environment. On Performance analysis of the proposed model using test data of the model through confusion matrix provides model accuracy of 98.4% which is found to be high compared other conventional architecture.