Prediction of Specific Language Impairment in Children using Cepstral Domain Coefficients

Specific Language Impairment (SLI) is a language disorder which prevents children from achieving language skills. Children that are affected with SLI are late to speak and may not generate any words until they are two years old. Children affected by SLI have difficulty constructing and comprehending coherent sentences. Most of the time, the effects caused by SLI last far beyond adolescence. Furthermore, due to absence of a clear cause of the illness, SLI is likely to go unnoticed by the majority of parents and teachers. This implies that automated identification of children with SLI is required. This paper presents an automated method of SLI detection with speech utterances with a relatively simpler and lighter machine learning model. The fundamental aim of this article is to develop a speaker-independent technique for SLI identification based on spoken utterances that is accurate and efficient using the minimum number of features. Two models are created that combine Mel-frequency cepstral coefficients (MFCC) with Neural network and Linear prediction coefficients (LPC) with Multi-layer Perceptron (MLP), respectively. These utterences are collected from SLI-diagnosed and healthy children’s speech samples. Overall results show that MFCC features when combined with MLP give the best performance. The best results acquired during model training, in comparison to accuracy rates of the aforementioned techniques, demonstrates the proposed method’s efficiency and efficacy.