Title : Vaani 2.0: An AI-powered vocal biomarker framework for preliminary Parkinson’s disease risk screening
Abstract:
Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by dopaminergic neuronal degeneration and motor and non-motor manifestations, including clinically relevant abnormalities in speech production. Conventional assessment remains primarily dependent on neurological examination and specialist evaluation, creating accessibility and scalability challenges for preliminary population-level screening. VAANI 2.0 is a full-stack artificial intelligence platform developed to investigate the feasibility of vocal biomarkers as a rapid, non invasive approach for preliminary PD risk assessment using short voice recordings. The proposed system combines browser-based audio acquisition with digital signal processing, acoustic feature extraction, and machine learning classification. Voice recordings are processed using the Librosa framework, from which 40 Mel-Frequency Cepstral Coefficient (MFCC) means and 40 MFCC standard deviations are extracted to form an 80-dimensional feature representation. A Random Forest classifier is then used to distinguish between healthy and Parkinsonian speech patterns and to generate a screening-oriented prediction and confidence estimate. In an evaluation comprising 73 voice recordings, with 15 samples reserved for testing, the model achieved an accuracy of 80.0%, precision of 80.0%, sensitivity of 66.7%, F1-score of 72.7%, and receiver operating characteristic area under the curve (ROC–AUC) of 88.9%. These findings demonstrate the feasibility of extracting discriminative information from vocal signals for preliminary PD risk screening while also highlighting the limitations imposed by the small dataset and test cohort. VAANI 2.0 is therefore positioned as a proof-of-concept clinical decisions upport and screening platform rather than a diagnostic system. The study provides a foundation for future investigation using larger, demographically diverse, clinically validated datasets and longitudinal evaluation.
