A Comparative Study of Lightweight CNNs for Dysarthria Classification

Dysarthria, a motor speech disorder that impacts speech clarity, creates a significant barrier to effective communication. This paper compares four lightweight Convolutional Neural Networks (CNNs), MobileNetV3Small, ShuffleNet, SqueezeNet, and EfficientNetB0, for classifying dysarthric speech using spectrogram images. These models are designed for deployment on devices with limited processing capabilities, where rapid inference is critical. The study considers classification accuracy, per-epoch training time, and number of parameters to assess both performance and the computational demands during model development. Results indicate that EfficientNetB0 achieves the highest accuracy (95.42%) but requires a substantially longer training time per epoch (265 s). This high training time implies that EfficientNetB0 is best utilized in settings where model training can be performed on more powerful hardware ahead of deployment, rather than in scenarios demanding frequent on-device retraining. In contrast, MobileNetV3Small processes data in less training time (28 s per epoch) albeit with lower accuracy (90.90%), while ShuffleNet offers a strong balance with 95.05% accuracy and moderate training time (55 s per epoch). Although SqueezeNet has the fewest parameters, its performance exhibits a higher standard deviation across cross-validation folds and a longer training time (420 s per epoch), indicating challenges in maintaining consistent results. These quantitative insights support informed model selection for dysarthric speech recognition (DSR) based on specific priorities, contributing to enhanced accessibility and assistive communication for individuals with speech impairments.