Voxidria

Project Story — Voxidria

What inspired us

Parkinson’s can affect speech early, but it’s not always obvious when changes are meaningful. We wanted to build a simple platform that turns short, structured voice tasks into an understandable risk indicator (not a diagnosis) that can encourage earlier, more informed conversations with professionals.

What we built

Voxidria is a full-stack web platform that:

Collects short voice recordings from guided speech tasks
Runs a machine learning model trained on speech-based datasets to produce a 0–100 risk score
Uses the Gemini API to explain the result in plain language (what it suggests, what it doesn’t mean, and practical next steps)
Lets users securely log in and view history over time

Tech stack (what we used)

Auth0 for secure authentication and account management
Supabase for backend storage and database (user data + recording metadata/history)
Gemini API for user-friendly explanations and educational guidance
Machine Learning model trained on datasets consisting of speech/pitch-related features for inference

How we built it (high-level flow)

User logs in (Auth0) to keep data private and tied to the right account.
Frontend guides voice tasks and uploads recordings to the backend.
Supabase stores user session-linked data and recording metadata/history.
The backend preprocesses audio, extracts key features, and runs ML inference to generate a risk score.
The platform sends structured model output to Gemini, which returns a clear explanation with disclaimers and suggested next steps.

Challenges we faced

Audio quality variance (background noise, mic differences, clipping) made consistent inputs harder, so we focused on structured tasks and preprocessing checks.
Health-related responsibility required careful wording and UX so the tool never implies diagnosis—only a risk signal.
End-to-end integration (Auth0 + Supabase + ML + Gemini) meant coordinating many moving parts and ensuring secure, reliable data flow.

What we learned

In health-adjacent projects, trust and clarity matter as much as the model.
LLMs are most helpful when given structured outputs and strict boundaries.
A strong demo needs a clean pipeline: record → validate → predict → explain → next step.