ORG-RGRU: An automated diagnosed model for multiple diseases by heuristically based optimized deep learning using speech/voice signal

Abstract

Human ailments create an impact in altering the significant metabolism activities of the body irrespective of various parts. Dysarthric speech is commonly known as Parkinson's disease (PD), where most of people suffer from voice impairment. Moreover, the voice analysis could help the clinicians to detect the disease effectively. Since it affects with neurological activities, the existing models are though render better results, but it cannot be able to meet up the desired outcomes. While diagnosing the disease by the clinicians, the abnormality variations in the voice signal become complicated. Due to the occurrence of unwanted interpretations, it affects the quality of the speech/voice signal. In order to meet this prerequisite, a novel diagnosing method is proposed. The proposed model implements the three-stage classification framework termed as Optimized ResNet and GoogleNet and Radial basis function-Gated Recurrent Unit (ORG-RGRU). The signals are garnered and decomposed using Empirical Wavelet Transform (EWT). The decomposed signal is given to classification in three different ways. Firstly, the decomposed signal is fed into Short-Time Fourier transform (STFT) features that are given into ORG-RGRU, which yields one classified output. Secondly, the relevant features are retrieved by adopting Mel-Frequency Cepstral Coefficients (MFCC), Cepstral and Spectral features, principle speech features, and pitch features (zero frequency response filter). Then, the weighted features are obtained with the optimal weight by the Adaptive Controlling Parameters-based African Vultures Optimization Algorithm (ACP-AVOA). Thus, the optimal features are fed into the Optimized Radial basis function-Gated Recurrent Unit (ORGRU) to exhibit the second classified outcome. Thirdly, the extracted STFT features are given into ResNet and GoogleNet, where the deep features are attained that are given to ORGRU. In order to return the optimal value, the hyper parameters in ResNet, GoogleNet, RBF and GRU are tuned optimally by ACP-AVOA. The high ranking is taken for determining the final classified results. The validation takes place among the developed model and existing conventional approaches. The main finding of the developed model shows 95% and 91% regarding accuracy and Matthews Correlation Coefficient (MCC).