VoiceCircle

Inspiration

AI can now clone a person's voice from a few seconds of audio. Scammers use it to phone grandparents as a "grandchild in trouble" who needs money right now. Older people are told to "be careful", but nobody lets them practice, and nobody notices when a call goes wrong. I wanted to build the family's safety net: practice before the scam, a check during it, and a gentle watch afterwards.

What it does

VoiceCircle lets a family form a private circle around an older relative.

• Practice calls: with the family member's consent, the app phones the elder with a realistic scam script in that relative's cloned voice. Every call ends by explaining it was practice, and the elder gets a score and feedback on whether she verified the caller.
• Check a call: upload a suspicious voicemail or recording. VoiceCircle reports whether it sounds AI-generated and whether it matches the real person, then tells you what to do next.
• Hugh, the daily companion: a friendly AI phones the elder each day for a short chat and sends the family a summary, so changes in mood or routine are noticed early.
• Alerts: anything worrying becomes an alert the family can review and mark as handled. Elders only need an ordinary phone. There is no app to install.

How we built it

• Backend: FastAPI (Python) with async SQLAlchemy, deployed on Render, and a scheduler for timed practice and check-in calls.
• Frontend: React and TypeScript with Vite, deployed on Vercel.
• Phone calls: Twilio Voice with speech recognition for the live back-and-forth.
• Voices: Fish Audio for consent-based voice cloning and speech.
• Detection: Resemble AI for AI-voice detection and speaker matching.
• AI conversation: an LLM served by Groq plays the caller and Hugh.
• Data and auth: Supabase for sign-in, Postgres and private audio storage.
• A built-in demo mode simulates calls, so the whole product can be tried without real phone numbers.

Challenges we ran into

• Making phone calls feel like a conversation: Twilio only hears speech while it is listening, so we nested the prompts inside the listening step to let people answer naturally.
• Staying inside strict timing limits: a phone webhook must answer within seconds, so slow AI responses fall back to a scripted line instead of leaving silence.
• Consent and privacy: voices are only cloned with explicit consent, are private to the circle, and can be deleted at any time.
• Connecting several providers (auth, storage, telephony, voice and AI) and handling their free-tier limits and failure cases gracefully.

Accomplishments that we're proud of

• A working end-to-end loop: enroll a voice, run a practice call, score it, check a fake recording, and raise an alert.
• Calls that respond to what the elder actually says.
• A simulator mode that lets anyone try the full flow in a browser.
• A backend covered by 134 automated tests.

What we learned

Voice security is as much about human habits as technology. A family code word and the habit of calling back on a known number beat any detector, so the product trains those habits first and uses detection as a second line of defense. We also learned how much real-time voice depends on latency, fallbacks and consent design.

What's next for VoiceCircle

• Support for Nigerian and other languages and Pidgin, so more elders can practice in the language they actually use.
• WhatsApp voice-note checks, since that is where many fake voice messages arrive.
• Family code-word setup and reminders.
• Stronger detection and a pilot with real families.

Built With

• fastapi
• fish-audio
• groq
• postgresql
• python
• react
• render
• resemble-ai
• sqlalchemy
• supabase
• twilio
• typescript
• vercel
• vite