Abstract
Automatic Speech Recognition (ASR) is a tough task, with the existence of related noise and high unpredictability in a speech presenting the most severe problems. Especially with regard to the noise of speech impairments, whether due to disability or mispronunciation in children. Extraction of noise-resistant features to compensate for speech degradation due to noise impact has remained a difficult challenge in the last few years. This research investigated the impact of different wavelet generations for extracting speech features, then test the produced dataset from each technique with two types of deep learning techniques: deep long short-term memory (LSTM) and hyper deep learning model convolutional neural network with long short-term memory (CNN-LSTM). The result shows that the deep long short-term memory of MFCC has reached 93% as an accuracy while in the hyper deep learning model of CNN-LSTM the accuracy of MFCC was 91%, as the highest recorded accuracy which proves that MFCC would be the best feature extraction technique for our developed dataset.
