VoiceCare AI

Voice-first AI triage for rural health workers — describe symptoms by voice in local languages, get instant risk screening for maternal, pediatric & malnutrition emergencies, even offline.

Comment

Inspiration

In rural India, over 900,000 ASHA (Accredited Social Health Activist) workers serve as the frontline of healthcare — yet they operate with minimal tools, low literacy support, and unreliable internet connectivity. Every year, preventable maternal and pediatric deaths occur simply because high-risk symptoms weren't flagged early enough for hospital referral.

We asked: What if an AI assistant could listen to a health worker describe symptoms in their own language and instantly flag life-threatening conditions?

What it does

VoiceCare AI is a voice-first health triage assistant that enables rural health workers to:

  • 🗣️ Speak symptoms naturally in English or Hindi — no typing required
  • 🧠 Get instant AI-powered risk assessment for maternal (pre-eclampsia, hemorrhage), pediatric (fever, respiratory distress), and malnutrition conditions
  • 🚨 Receive clear action steps — from home monitoring to immediate hospital referral with ambulance instructions
  • 📡 Work completely offline as a PWA — critical for areas with poor connectivity
  • 📋 Track patient history locally for follow-up visits

How we built it

We built VoiceCare AI as a lightweight Progressive Web App using vanilla HTML, CSS, and JavaScript — intentionally avoiding heavy frameworks to ensure fast loading on low-end devices common in rural areas.

  • Voice Engine: Web Speech API for real-time speech-to-text (supporting Hindi and English) with live waveform visualization
  • AI Triage Engine: A keyword-based NLP system with 25+ symptom patterns, severity-weighted scoring, and context-aware modifiers (pregnancy status, patient age)
  • Risk Scoring: A 0-100 risk score algorithm that factors symptom severity, quantity, and patient context to classify into LOW / MEDIUM / HIGH risk levels
  • Offline-first PWA: Service worker caches all assets for full offline functionality
  • Mobile-first UI: Dark gradient theme with large touch targets (56px+) designed for field use under sunlight

Challenges we faced

  • Dialect variation: Hindi medical terminology varies greatly across regions. We addressed this by including multiple keyword variants for each symptom
  • Offline constraint: The entire triage engine had to run client-side with zero API calls, pushing us to build a comprehensive local knowledge base
  • Low-literacy UX: Designing for users who may not be comfortable reading — voice input/output and large visual indicators were our solution
  • Balancing sensitivity vs specificity: The triage engine needed to err on the side of caution (flagging potential emergencies) without creating alarm fatigue

What we learned

  • Voice-first interfaces dramatically lower the barrier to technology adoption in low-resource settings
  • A well-designed keyword NLP system can achieve surprisingly effective symptom detection without requiring cloud-based LLMs
  • PWA technology is a game-changer for healthcare in connectivity-challenged areas

What's next for VoiceCare AI

  • 🌐 Adding more regional languages (Tamil, Telugu, Bengali, Marathi)
  • 🤖 Integration with on-device LLMs (Gemini Nano) for conversational symptom gathering
  • 📊 Dashboard for PHC (Primary Health Centre) supervisors to monitor field data
  • 🔗 Integration with government health databases (RCH portal, HMIS)

Built With

Share this project:

Updates