Speech impairment detection in children using time frequency features of speech and deep learning techniques.

Speech impairment in children is nowadays occurring more commonly than in earlier days. It is necessary to detect the impairment in speech as early as possible to provide therapeutic treatments. In this work, a novel method has been suggested for speech impairment detection. First, speech signals are collected from children, including vowels, consonants, and different syllables. The speech signals are then processed using gammatone filter bank-based cepstral coefficients (GFCC) and discrete wavelet transform (DWT) for time-frequency feature extraction. DWT is a time-frequency feature that is extracted with the Haar wavelet. The proposed work uses a 5-fold cross-validation method for implementation. The features are then given to various learning technologies for designing the training speech impairment detection module. The learning technologies used are shallow learning (multilayer perceptron (MLP) and support vector machine (SVM)), shallow ensemble learning (gradient boosting (GBoost), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), light gradient boosting machine (lightGBM), Stacking) and deep learning (Convolutional long short term memory (ConvLSTMs), bidirectional LSTM (BiLSTMs), convolutional neural network LSTM (CNN-LSTM), gated recurrent units CNN (GRU-CNN)) methods. The training module is then tested using speech signals that were not given in training. The test results of the methods have been analysed to select an optimal method. It is observed that the highest accuracy of 99.67% is obtained using the GRU-CNN method with DWT features. The contribution of the proposed method is that the DWT features with the GRU-CNN model have not been suggested for speech impairment detection until now with this accuracy. Yet another contribution is that all speech signals combined are given as input in one module, unlike existing methods that are designed with a separate module for each type of speech. Based on the findings, it is feasible to develop a computational framework for detecting speech impairment based on time frequency representations of speech signals along with deep learning methods.