Screening and analysis of specific language impairment in young children by analyzing the textures of speech signal.

A child having a delayed development in language skills without any reason is known to be suffering from specific language impairment (SLI). Unfortunately, almost 7% kindergarten children are reported with SLI in their childhood. The SLI could be treated if identified at an early stage, but diagnosing SLI at early stage is challenging. In this article, we propose a machine learning based system to screen the SLI speech by analyzing the texture of the speech utterances. The texture of speech signals is extracted from the popular time-frequency representation called spectrograms. These spectrogram acts like a texture image and the textural features to capture the change in audio quality such as Haralick's feature and local binary patterns (LBPs) are extracted from these textural images. The experiments are performed on 4214 utterances taken from 44 healthy and 54 SLI speakers. Experimental results with 10-fold cross validation, indicates that a very good accuracy up to 97.41% is obtained when only 14 dimensional Haralick's feature is used. The accuracy is slightly boosted up to 99% when the 59-dimensional LBPs are amalgamated with Haralick's features. The sensitivity and specificity of the whole system is up to 98.96% and 99.20% respectively. The proposed method is gender and speaker independent and invariant to examination conditions.