Automatic recognition of disordered children’s speech signal in dyadic interaction using deep learning models

Children suffering with spontaneous speech impairment or inappropriate communication abilities like disordered speech or delayed speech face challenges when involved in conversations. One of the motivating reasons for this work is to use the potential of deep learning models along with an effective feature extractor to automate the detection of specific language impairment (SLI) in children. Clinicians or speech pathologists use standard assessment tools that are time consuming as well as prone to various behavioural factors which can compromise the timely identification of the SLI in children. Moreover, the scarcity of annotated disordered children’s speech adds to the complexity of training the reliable SLI detection model. The recent work focuses mainly on the utterance of vowels, scanning the acoustic features or the texture of the children’s speech signals to detect the SLI. This work seeks to evaluate different components of children’s speech like vowels, consonants and sentences to diagnose healthy and disordered speech. Speech samples are collected from Indian children in the age-group of 5-15 years speaking a secondary language English. The proposed method makes use of a combination of mel frequency cepstral co-efficients and i-vectors as a feature vector to identify SLI and distinguish it from mispronunciations due to second language usage. Moreover, analysis of variance (ANOVA) test has been implemented to choose the most significant MFCC and i-vector features. Finally, the selected features are given as input to pretrained models like VGG-16, MobileNet-v2, ResNet-50 and ResNet-101. Eventually, we study and evaluate the effect of parameters like age and model used on different parameters of speech like vowels, consonants, 3-word, 4-word and 5-word sentences in the dataset to diagnose healthy and disordered speech samples. 5-fold cross validation (CV) has been used to compensate for the limited size of dataset and achieve robust results. The experimental results show that with the proposed implementation method highest accuracy of 98.70% can be achieved on the vowels component in identifying the disordered children speech signals using MobileNet-v2 model.