Abstract:
Specific Language Impairment is a language disorder which prevents children from achieving language skills. Children that are affected with SLI are frequently late to speak and may not generate any words until they are two years old. Children affected by SLI have difficulty constructing coherent sentences. Moreover, they have difficulty in comprehending them. Most of the time, the effects caused by SLI last far beyond adolescence. From the age of two until the age of six, a child’s speech develops. Hence, around the age of 3 or 4 years is the best time for SLI diagnosis. 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. The previously reported research on this dataset has shown good accuracies, but they were achieved using complex classifiers. Moreover, to achieve that accuracy, they utilized a relatively large set of features. In addition to that, this study is based on the entire range of the analysed child’s utterances, instead of just vowels. The fundamental goal of this work is to create 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 MFCC characteristics with MLP and LPC characteristics with MLP, respectively. These characteristics are collected from SLIdiagnosed and healthy children’s speech samples. The speech utterances used in this work were acquired from a database established by the LANNA research group. Overall results show that MFCC features when combined with MultiLayer Perceptron 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.
