Specific language impairment detection through voice analysis

Abstract

Specific Language Impairment is a communication disorder regarding the mastery of language and conversation that impacts children. The system proposed aims to provide an alternative diagnosis method that does not rely on specific assessment tools. The system will accept a voice sample from the child and then detect indicators that differentiate individuals with specific language impairment from that voice sample. These indicators were based on the timbre and pitch characteristics of sound. Three different feature spaces are calculated, followed by derived features, with three different classifiers to determine the most accurate combination. The three feature spaces are Chroma, Mel-frequency cepstral coefficients (MFCC), and Tonnetz and the three classifiers are Support Vector Machines, Random Forest and a Recurrent Neural Network. MFCC, representing the timbre characteristic, was found to be the most accurate feature vector across all classifiers and Random Forest being the most accurate classifier across all feature spaces. The most accurate combination found was the MFCC feature vector with the Random Forest classifier with an accuracy level of 99%. The MFCC feature vector has the most features that are extracted giving the reason for the high accuracy. However, this accuracy decreases when the recorded word is three syllables or longer. The system proposed has proven to be a valid method that can detect SLI.