Scalogram based performance comparison of deep learning architectures for dysarthric speech detection

Abstract

Dysarthria, a speech disorder commonly associated with neurological conditions, poses challenges in early detection and accurate diagnosis. This study addresses these challenges by implementing preprocessing steps, such as noise reduction and normalization, to enhance the quality of raw speech signals and extract relevant features. Scalogram images generated through wavelet transform effectively capture the time-frequency characteristics of the speech signal, offering a visual representation of the spectral content over time and providing valuable insights into speech abnormalities related to dysarthria. Fine-tuned deep learning models, including pre-trained convolutional neural network (CNN) architectures like VGG19, DenseNet-121, Xception, and a modified InceptionV3, were optimized with specific hyperparameters using training and validation sets. Transfer learning enables these models to adapt features from general image classification tasks to classify dysarthric speech signals better. The study evaluates the models using two public datasets TORGO and UA-Speech and a third dataset collected by the authors and verified by medical practitioners. The results reveal that the CNN models achieve an accuracy (acc) range of 90% to 99%, an F1-score range of 0.95 to 0.99, and a recall range of 0.96 to 0.99, outperforming traditional methods in dysarthria detection. These findings highlight the effectiveness of the proposed approach, leveraging deep learning and scalogram images to advance early diagnosis and healthcare outcomes for individuals with dysarthria.