VAANI 2.0

Home_Page
Scrolldown_from_Home Page
Analyzing
Result_1
Result_2
Results_Proof_From_Backend

A son and his father used to talk every evening. No matter how busy life became, they always found time for a conversation. Sometimes it was about school, sometimes about work, and sometimes about nothing important at all. The conversations themselves were what mattered. Then one day, the son noticed something strange. His father's voice sounded different. Words were softer. Sentences became slower. Sometimes there were unusual pauses. Everyone assumed it was stress, age, or fatigue. Life continued. Months passed. The pauses became more frequent. The voice became weaker. Simple tasks became harder. Eventually, the family received a diagnosis that changed everything. Parkinson's Disease. By the time the diagnosis was confirmed, significant neurological damage had already occurred. The symptoms that everyone could see were only the final chapter of a disease that had been silently progressing for years. The son could not stop wondering: "What if we had known earlier?" "What if the voice changes had been recognized as a warning sign?" "What if there was a simple way for anyone, anywhere, to screen themselves before the disease progressed?"

This story is not unique.

Millions of families around the world experience the same reality. More than 10 million people worldwide live with Parkinson's Disease, and research has shown that vocal changes can appear years before many of the more recognizable motor symptoms. That question became the inspiration behind VAANI.

VAANI 2.0 is an AI powered Parkinson's vocal biomarker screening platform.

Users upload a voice recording, and the system analyzes acoustic biomarkers associated with Parkinsonian speech. Using machine learning and digital signal processing techniques, VAANI evaluates features such as:

Jitter, Shimmer, Harmonics to Noise Ratio (HNR), Spectral characteristics, Voice stability patterns

The platform then generates:

A risk score, Model confidence, Acoustic feature analysis, An AI generated screening report, Downloadable PDF results

VAANI is designed as an accessible early screening tool that can help identify potential warning signs and encourage further clinical evaluation.

How I built it

VAANI 2.0 builds upon our previous project, VAANI 1.0.

VAANI 1.0 was developed as a complete ESP32 based hardware screening device, for the edge impulse hackathon, that demonstrated the concept of voice based Amyotrophic lateral sclerosis (ALS)'s detection. While it proved the feasibility of the idea, I wanted to create a more scalable and intelligent platform.

VAANI 2.0 was built as a full stack AI powered web platform. We developed a modern frontend using React, TypeScript, and Tailwind CSS to create an intuitive dashboard where users can upload voice recordings, view acoustic biomarkers, and download detailed reports. On the backend, we used Python and Flask to process audio recordings and perform feature extraction. For the machine learning component, we used Librosa, NumPy, Scikit learn, and a Random Forest model trained on healthy and Parkinsonian voice recordings. The system analyzes MFCC features along with vocal biomarkers such as jitter, shimmer, and HNR to generate risk assessments. We also integrated PDF report generation so users can save and share their results. By moving from a standalone hardware device to an AI powered platform, we were able to significantly reduce costs and make early Parkinson's screening available to a much larger population.

Challenges I ran into

One of my biggest challenge was dataset quality.

During testing, I discovered that some recordings contained variations in speaking style, pauses, and fluency. This created a risk that the model might learn differences in recording behavior rather than true Parkinsonian characteristics.

Here's the link to the dataset I used: https://www.kaggle.com/datasets/asthamishra96/parkinson-multi-model-dataset-2-0?

Another challenge was the limited amount of available Parkinson's voice data. Training a robust medical AI system requires significantly larger and more diverse datasets than those readily available. Integrating machine learning with a professional looking frontend also presented difficulties. We had to connect audio upload pipelines, backend processing, real time inference, dynamic dashboards, and PDF generation into a single seamless experience.

Finally, balancing accuracy and interpretability was challenging. We wanted users to understand why a prediction was made rather than simply receiving a number.

Accomplishments that I am proud of that VAANI evolved from a hardware prototype into a complete AI powered platform. Building VAANI 2.0 also helped me understand how mathematics powers modern artificial intelligence. I learned how audio signals can be represented as numerical waveforms and transformed into meaningful features using mathematical techniques such as Mel Frequency Cepstral Coefficients (MFCCs). The project involved concepts from statistics, including means, standard deviations, probability, and model confidence scores. I also explored how machine learning algorithms use mathematical patterns in data to make predictions. Working on VAANI showed me that mathematics is not just something learned in a classroom, but a powerful tool that can be used to solve real world healthcare challenges and potentially improve lives.

here's a paper that i wrote on vaani: https://zenodo.org/records/17985539

This project taught me far more than just programming and machine learning. While building VAANI 2.0, I learned about audio signal processing, feature extraction using Librosa, full stack development, API integration, model evaluation, and the ethical responsibilities involved in developing AI for healthcare. However, the most valuable lesson was understanding the human side of technology. I realized that behind every voice recording is a real person, a real family, and a real story. Throughout development, I learned how important dataset quality and bias are, especially when people's health may be affected by the results. Most importantly, I learned that technology should not only be innovative but also meaningful. Building VAANI helped me understand that the true purpose of engineering is not simply to create technology, but to use it to improve lives and help people when they need it most.

I also learned that creating AI for healthcare is not only a technical challenge but a human one. Every percentage point of accuracy can represent real people, real families, and real lives.

What's next for VAANI 2.0

VAANI 2.0 is only the beginning, my long term vision is not simply to detect Parkinson's Disease earlier, my goal is to help defeat it. Most importantly, I aim to expand from early detection toward supporting researchers and clinicians in the search for better treatments and, one day, a cure for Parkinson's Disease.

If a voice can provide the first warning sign, then perhaps technology can help ensure that families never have to ask the question:

"What if we had known sooner?"