Speech Analysis with Deep Learning to Determine Speech Therapy for Learning Difficulties

Abstract

The vocal tract movements involved in human speech makes the vocalisation of a complex array of coordinated and meaningful acoustic utterances possible. At the same time, it is hypothesized that related cognitive disorders can potentially interfere with the neurological, pre-articulatory and fine motor controls required for these fine movements. By leveraging the cognitive complexity of speech production, it is possible to detect a range of different disorders. Computer screening systems can be considered as an efficient approach for the early diagnosis and screening of voice disorders. For achieving the highest detection rate possible, a hybrid machine learning-based approach is proposed by combining Deep Learning with AdaBoost classifier. First, a set of acoustic features will be extracted using traditional features associated with the presence of autism, such as fundamental frequency descriptors. Then, a deep learning framework will be utilized for extracting additional acoustic contextual descriptors not definable using traditional feature extraction methods. Finally, the most informative features will be selected using a minimal-redundancy maximal-relevance feature selection approach with an AdaBoost classifier analysing all the selected features and informing the operator regarding the patient’s condition.