Digital Biomarkers for Laryngeal Health Detection: A Multi-Modal CNN Approach Integrating 2D-Bispectrum and High-Dimensional MFCCs

Abstract

Identifying speech dysphonia remains a constant diagnostic difficulty because of the non-linear character of glottal turbulence. Critical phase coupling properties intrinsic in pathological vocal fold vibrations are sometimes missed by conventional second-order power spectra analyses. To tackle this, we suggest an inventive Dual-Branch Convolutional Neural Network (DB-CNN) architecture meant to simultaneously record high-dimensional spectral features and higher order statistics. The model processes Mel-Frequency Cepstral Coefficients (MFCC) and Bispectrum Magnitude Maps separately using a Late Fusion method based on a Directed Acyclic Graph (DAG). A balanced dataset $(\mathbf{N}=\mathbf{8 1 0})$ experiment shows a considerable evolutionary performance trend. Though a single-branch MFCC-13 model only reached 0.816% accuracy, adding third order statistics and increasing spectral resolution (MFCC 14-26) significantly boosted diagnostic effectiveness. With Sensitivity of 86.89% and Specificity of 90.55%, the suggested DB-CNN achieved a maximum Accuracy of 89.36%. With a Matthews Correlation Coefficient (MCC) of 0.7623, the model is clearly reliable and a strong diagnostic tool. These findings indicate that the interaction between high-order MFCCs and non-linear Bispectrum analysis produces a very efficient digital biomarker for dysphonia identification, therefore surpassing baseline fusion techniques dramatically.