CrisisVoice

Inspiration

Every day, crisis hotline responders are the last line of defense for people in emotional freefall. But even the most experienced professionals can miss subtle cues in a fast-paced, high-stress call. We were driven by one goal:
What if we could give every hotline operator an AI-powered co-pilot — one that listens, analyzes, and supports in real time?

CrisisVoice was born from this mission: not to replace human empathy, but to enhance it with insight.

What It Does

CrisisVoice transforms crisis hotline conversations — whether text transcripts or audio recordings — into actionable mental health insights.

Using a custom-trained NLP classifier model, the system detects:

Risk factors (e.g., hopelessness, social withdrawal)
Protective factors (e.g., mention of family, reasons for living)
Emotional tone and urgency
Assigns an overall suicide risk score
Generates explainable insights and recommended next steps

It’s like having an AI-powered observer that highlights what matters, so responders can act faster, with more confidence.

How We Built It

Frontend: Built in React and Tailwind CSS for a clean, accessible, and responsive UI
Backend: FastAPI server powering secure endpoints for model inference and file processing
Model: Trained NLP classifier to detect mental health risk indicators, informed by psychology research such as the Interpersonal Theory of Suicide
Speech-to-Text: Integrated Whisper to convert .mp3 crisis call recordings into analyzable transcripts
LLM Integration: Used Gemini to translate raw model outputs into understandable, human-friendly summaries for hotline operators

All data is handled with strict privacy — no personal info is stored, and files are processed in secure memory.

Challenges We Ran Into

Designing a sensitive, non-triggering interface that respects the emotional weight of crisis data
Getting multi-GPU training synchronization working for faster NLP model iteration
Translating raw probabilities into meaningful mental health insights
Handling low-resource emotional tone detection within our 24-hour window

Accomplishments We're Proud Of

Built a fully functional and demo-ready platform in under 24 hours
Seamlessly integrated speech-to-text, NLP, LLMs, and real-time scoring
Created a tool that could make a real difference in life-and-death conversations
Followed ethical AI practices, especially around sensitive data handling and explainability

What We Learned

How to turn a bold idea into a working MVP under extreme time constraints
The importance of human-centered AI in mental health
How to align psychological theory with model architecture for real-world applicability
Mastered key tech skills in end-to-end ML pipeline design, UI/UX for crisis response, and LLM-driven interpretation

What's Next for CrisisVoice

We believe CrisisVoice can be a real tool for social good — and we're just getting started.

Future directions:

Real-time streaming call analysis (e.g., Twilio integration)
Multilingual support for global crisis services
Fine-tuned models on verified, clinical-grade datasets
Integration with EHR systems and responder dashboards
Partnerships with nonprofits, campus mental health centers, and national suicide prevention hotlines

CrisisVoice is designed to be quiet, respectful, and powerful — a trusted digital partner for the people who save lives.