Fatigue Classification Using Chalder Fatigue Scale Based Vocal Biomarkers, VGGish Deep Representations, Prosodic Features and Meta Model Fusion

Abstract

Voice is a compelling, non-invasive biomarker for fatigue, yet much of the field's reported success rests on evaluation protocols that quietly permit information leakage and collapse once that leakage is removed. This paper reframes the question, asking not how high accuracy can be pushed but how much fatigue signal the human voice genuinely carries once every avenue for leakage is closed. To this end we develop and evaluate a hybrid voice-based fatigue classification framework built around honest, reproducible measurement. The framework integrates transfer-learned deep visual embeddings with interpretable prosodic and cepstral descriptors, classifies the resulting hybrid representation through five calibrated classifier families (logistic regression, histogram gradient boosting, support vector machines, multilayer perceptron, and linear discriminant analysis), and combines them through out-of-fold, reliability-weighted fusion. An integrity-audit procedure comprising a determinism check, label-permutation testing, calibration assessment, and multi-seed stability analysis is embedded within training. Evaluation on two Turkish corpora with one recording per speaker, a 468-recording reading task and a 462-recording spontaneous-speech task, is strictly cross-speaker. The deployed fusion achieves out-of-fold area under the curve of 0.603 and 0.607, multi-seed stability of 0.595 and 0.596, and held-out area under the curve of 0.569 and 0.631. Four independent estimates converge near 0.60, locating a ceiling imposed by the data rather than the model. The contribution is methodological: a leakage-audited reference point and a transferable, self-auditing protocol on which larger and better-labeled studies can build.