Can Voice Reveal Neurodegeneration? A Theory and Prototype of an Offline AI Based Screening Device

Neurodegenerative diseases such as Amyotrophic Lateral Sclerosis (ALS) and Parkinson’s disease are often diagnosed at advanced stages, when significant and irreversible neuronal damage has already occurred. Existing diagnostic techniques, including MRI, PET scans, and neurological examinations, but one downside is that these methods are expensive, time consuming, and largely inaccessible to rural and low resource populations. As a result, early screening remains a major global challenge. This paper presents Vaani, a low cost, portable, and fully offline voice based screening system designed to identify early speech impairments associated with neurodegenerative disorders. The system is based on the observation that diseases affecting motor neurons disrupt fine control of the vocal apparatus, leading to measurable abnormalities in speech such as vocal tremor, pitch instability, amplitude variation, and formant distortion. These changes are captured using Mel Frequency Cepstral Coefficients (MFCCs), which model the short-term spectral envelope of speech. Mathematically, sudden frequency drops and spikes, common in ALS affected speech due to impaired motor coordination, manifest as abrupt variations in the short time Fourier spectrum. These irregularities propagate into MFCC feature space as high temporal variance across adjacent frames, which can be learned by lightweight temporal models. Vaani employs a Conv1D neural network to analyse these temporal MFCC patterns and classify speech into three risk categories: healthy, Parkinson’s risk, and ALS risk. The entire pipeline runs on an ESP32 microcontroller using an INMP441 MEMS microphone, with inference deployed via TensorFlow Lite, achieving screening results in under six seconds without internet connectivity. Training was conducted using publicly available datasets (Kaggle and Synapse), with additional internally collected samples to improve robustness. Designed using beginner accessible tools such as Edge Impulse Studio and mobile based CAD platforms, Vaani demonstrates that meaningful neurological screening systems can be built with minimal cost and technical barriers.