Comparative analysis of different time-frequency image representations for the detection and severity classification of dysarthric speech using deep learning

Dysarthria, a speech disorder resulting from neurological damage, presents significant challenges in clinical diagnosis and assessment. Traditional methods of dysarthria detection are often time-consuming and require expert interpretation. This study analyzes various time-frequency image representations of TORGO dysarthric speech to facilitate the automatic detection and classification of dysarthria severity through deep convolutional neural networks (DCNN). The dysarthria detection problem was approached in experiment E1, a binary classification task involving dysarthria and a healthy control class. Experiment E2 employs a multiclass classification method, categorizing data into very low, low, medium, and healthy classes. The analysis of time- frequency image representations of speech features is presented in two forms: standard-form images, including cepstrogram and spectrogram, and compact-form images, such as cochleagram and mel-scalogram. The highest ranked feature is benchmarked with existing work for both dysarthria detection and its severity classification. And this proposed work analyzes the frequency behavior of time-frequency image representations by bifurcating the standard-form images into two halves: one half representing low frequency and the other half representing high frequency. By this approach, the bifurcated standard-form of cepstrogram with low frequency outperforms all other features by achieving a validation accuracy of 99.53% for E1 and 97.85% for E2 and this surpasses existing benchmark work by 6% for E1 and by 1.7% for E2 in the TORGO dataset. The best ranked feature of E1 and E2 was applied to the noise-reduced UASpeech dataset and the DCNN model achieved 98.75% accuracy in E1 and 95.98% in E2, demonstrating its effectiveness on new dataset.