TSCE and CD-CATL Driven Framework for Robust and Real-Time Voice Disorder Detection and Classification

The creation of non-invasive methods for precise and non-invasive diagnosis of vocal disorders in clinical speech diagnostics is tremendously challenging owing to the tremendous variation in demographic, linguistic, and acoustic features. In this paper, a powerful deep learning-based system is proposed that is capable of identifying and classifying vocal fold defects using the Aachen Voice Pathology Database (AVPD) using Temporal Spectro-Context Encoding (TSCE) and Cross-Domain Context-Aware Transfer Learning (CD-CATL). The dataset contains 388 annotated high-quality speech samples that cover a wide range of conditions, such as paralysis, edema, nodules, and polyps. The data are time-corrected following Gammatone-based spectrotemporal decomposition with dynamic time warping and short-time Fourier transform in the preprocessing pipeline. The TSCE module maintains phonatory dynamics while encoding local and distant acoustic interactions by employing dilated convolutions and multi-head attention. The system is learned to acquire domain-invariant features while maintaining disease-specific representations by combining memory-augmented transformer streams with multi-scale convolutional attention in the CD-CATL architecture. The model performs better than baseline CNN and RNN models on all standard evaluation measures, with a sensitivity of 97.81%, specificity of 98.56%, and an accuracy of 98.89%. The system is appropriate for telehealth use with its real-time inference enabled by its low-latency optimized deployment with ONNX and TensorRT. The suggested approach seems to have the potential for providing clinically sound, scalable, and objective voice disorder screening for use across a range of low-resource health care environments.