Self-Organized Feature Engineering Model for Accurate Speech Disorder Classification

Dysarthria, spasmodic dysphonia and stuttering are still diagnosed mainly through perceptual assessment, which is subjective and difficult to scale. This study presents a lightweight and explainable feature engineering model that distinguishes these three disorders from normal speech. Its core is the expansion–contraction pattern (ExCoPat), a lattice-based textural descriptor whose traversal path is determined by the data of each signal block. ExCoPat and twenty statistical measures are computed from the raw signal and from eleven low-pass sub-bands of a db4 multilevel discrete wavelet transform, giving 10,176 features per clip. Iterative neighborhood component analysis and iterative ReliefF select the informative features, k-nearest neighbors and support vector machine classifiers generate the predictions, and iterative majority voting with greedy selection provides the final decision. On a curated dataset of 5,244 three-second clips, the model achieved 98.95% accuracy, 98.92% unweighted average recall and 98.94% F1-score under 10-fold cross-validation. Under the same protocol, an artificial neural network with three hidden layers trained on cepstral statistics and a shallow three-layer convolutional neural network trained on log-mel spectrograms reached only 93.97% and 89.84% accuracy, respectively. Removing the ExCoPat features reduced accuracy by 11.27 percentage points, confirming their central role. Feature extraction required 433 ms per clip on a CPU, and the extractor has no trainable parameters. These results suggest that a transparent handcrafted pipeline can serve as a practical tool for multi-disorder speech screening.