Advancing clinical group verification framework for screening child speech sound disorders using “text-independent” i-Vectors

I-Vectors are the current state-of-the-art feature representation in acoustic event identification tasks such as speaker recognition, language recognition, etc. They are referred to as identity vectors since they represent a unique quality of the speaker and have been useful for detecting adult speech pathologies . Speech Sound Disorders (SSDs) affect between 3% and 16% of US children and are difficult to detect due to the presence of developmental speech sound errors. We have been working to automate the process of speech screening. Our dataset consists of 29, single word recordings from 165, 3–6 year old children and was collected using an iOS application. Children were assigned to clinical groups using a percentage consonants correct growth curve model. Sixty-four children were classified as exhibiting an SSD, the rest as exhibiting normal speech acquisition. To achieve our purpose, we first introduced our clinical group verification framework using Gaussian Mixture Models. We extended the framework to screen the children’s speech based on single words using “text-dependent” i-Vectors, along with L2-logistic regression and Gaussian backend machine learning classifiers. This improved the algorithms’ accuracy. In the current study, we modified our offline post-processing algorithms within the framework, which provided excellent results for “text-independent” i-Vectors. Apart from the algorithms, we also present detailed visual analysis of the classifier score transformation, during the post-processing phase.