Bio Larynx: When Your Voice Knows Before You Do

Abstract

Neurological and vocal disorders such as Parkinson’s Disease, Alzheimer’s Disease, stroke related impairments, and vocal fold pathologies often go undetected in their early stages due to a reliance on subjective assessments, expensive imaging techniques, or invasive procedures. These challenges make it difficult to diagnose conditions early, monitor them effectively, and ensure accessible care, especially in settings with limited resources. To demonstrate the potential of artificial intelligence in this context, this study analyzes speech derived biomarkers from existing datasets to explore how AI can assist in identifying and understanding neurological and vocal disorders. Built upon publicly available datasets, the system integrates a complete analytical pipeline including data preparation, exploratory analysis, feature selection, machine learning based classification, and statistical interpretation. It employs both statistical methods and ensemble learning techniques to identify robust acoustic and prosodic features relevant across multiple disorders. A suite of classifiers ranging from logistic regression to gradient boosting and neural networks are trained and validated using stratified cross validation. Significant biomarkers are further examined using nonparametric tests and effect size estimation. Analysis on various parameters were carried out like Fundamental Frequency, Speech Duration, Pitch period Entropy, Detrended Fluctuation analysis and more. Unsupervised clustering and dimensionality reduction techniques are also applied to explore latent subgroups within patient populations. The system outputs interactive visualizations and auto generated reports, offering a transparent, reproducible, and scalable approach to voice based health diagnostics.