Classification Between Patients with Amyotrophic Lateral Sclerosis and Healthy Individuals Using Hypernasality in Speech: A Low Complexity Approach*

Hypernasality, commonly observed in dysarthric speech due to impaired velopharyngeal function, is a key charac-teristic of Amyotrophic Lateral Sclerosis (ALS). The increasing demand for lightweight models capable of operating on resource-constrained platforms, such as mobile devices or general-purpose computers, underscores the importance of developing efficient methods for ALS analysis. This study explores hypernasality as a potential indicator for ALS by utilizing nasal and non-nasal phonemes from healthy speech for ALS vs. Healthy Controls (COT) classification. We use HuBERT, a pre-trained speech representation model, alongside a Dense Neural Network (DNN) for classification. We also explore reducing model complexity by minimizing the number of dense layers in a DNN model and compare its performance with that of a higher-complexity DNN model. Experiments involving 57 ALS patients and 55 COT using Spontaneous Speech (SPON) and Diadochokinetic Rate (DIDK) tasks show that hypernasality can effectively distinguish ALS from COT. The HuBERT layer that provides the highest classification accuracy is selected based on results from Test Set 1 (30 ALS and 30 COT), and the performance of the low-complexity models for this layer is evaluated on Test Set 2 (27 ALS and 25 COT). The highest classification accuracy for the SPON task is 73.47%, using features from the second layer of the HuBERT model, while the DIDK task achieves 79.15% accuracy with features from the eleventh layer. Reducing model complexity leads to minimal average accuracy loss across various train-test conditions- 3.07% and 4.43% for SPON, and 5.44% and 5.90% for DIDK-while achieving substantial reductions in model parameters (14.71 % and 98.67%) and floating-point operations (FLOPs) (14.44% and 98.66%). Notably, the lowest complexity model, for DIDK, achieves 81.46% accuracy with just 1,538 parameters and 1,540 FLOPs, compared to 115,714 parameters and 115,200 FLOPs in the high-complexity model. For SPON, it achieves a maximum accuracy of 66.48%. This work demonstrates that hypernasality serves as an effective ALS indicator, and reduced model complexity provides a feasible trade-off between performance and resource efficiency.