Vox | Devpost

Vox

Inspiration

We all have moments when we need guidance—before a job interview, learning a new language, or just processing a stressful day. But coaches, tutors, and mentors aren't available 24/7.

What if your phone contacts included AI experts you could actually call anytime? Not chatbots with text bubbles, but voices you recognize—assistants that feel like real people in your pocket.

And the big one: traveling abroad and wishing you could speak the local language in your own voice? That sparked the real-time translator with voice cloning.

What it does

Vox turns AI into voice contacts you can call anytime.

AI Contacts: Interview coach Alice, Spanish tutor Carlos, wellness mentor Dr. Sam—each with unique voices and personalities
Push-to-Talk: Hold to speak, release to send. Natural voice conversation flow
Voice Cloning: Record 30 seconds of your voice, clone it, and use it across contacts
Real-Time Translator: Speak in English, hear yourself respond in Spanish, French, German, Japanese—30+ languages in YOUR voice
Multi-AI Support: Choose between Gemini, Claude, GPT-4, or DeepSeek per contact
Works Everywhere: Web, PWA, and Telegram Mini App

How we built it

Layer	Tech
Frontend	Next.js 16, React 19, TypeScript, Tailwind CSS
AI Brain	Google Cloud Vertex AI (Gemini 2.0), Claude, GPT-4, DeepSeek
Voice	ElevenLabs TTS, STT, Voice Cloning API
Backend	Firebase Auth, Firestore, Cloud Storage
Observability	Datadog LLM Observability, RUM, Detection Rules
Infrastructure	Upstash Redis (rate limiting), Stripe (payments)

The architecture prioritizes voice-first UX—minimal latency from speech to AI response to spoken reply.

Challenges we ran into

Voice latency: Getting under 2 seconds from user speech → AI response → TTS playback required aggressive optimization and streaming
Conversation context: Maintaining natural dialogue across voice turns without losing context
Voice cloning quality: Ensuring 30 seconds of audio produces a recognizable, natural-sounding clone
Multi-provider routing: Seamlessly switching between Gemini, Claude, and GPT-4 while maintaining consistent UX
Telegram Mini App constraints: Adapting the full experience to work within Telegram's webview limitations

Accomplishments that we're proud of

Voice cloning that actually works: Users hear themselves speaking languages they don't know
Sub-3-second response loop: Speak → AI thinks → Voice responds, feels like a real call
Enterprise-grade observability: Full Datadog LLM monitoring with detection rules that auto-create incidents
5 ready-to-use AI contacts: Each with distinct personality, voice, and purpose
Cross-platform: Same experience on web, mobile PWA, and Telegram

What we learned

Voice UX is fundamentally different from chat UX—timing and tone matter as much as content
LLM observability isn't optional at scale—you need to track latency, costs, and errors from day one
ElevenLabs voice cloning is surprisingly good with just 30 seconds of clean audio
Users form emotional connections with AI voices faster than with text-based assistants

What's next for Vox

Group Calls: Mix AI contacts with real friends—imagine a study group with your friend, an AI tutor, and GPT-4 brainstorming together
AI Music Creation: Contacts that compose and send personalized music to friends
Memory & Relationships: Each contact remembers your conversations, preferences, and builds a relationship over time—like a real friend who knows your history
Voice Avatars: Video avatars that lip-sync with the AI voice for video calls
Enterprise Version: Custom AI contacts for company onboarding, training, and support

Pitch Deck: https://vox-aicontact-fe0e3.web.app/pitch-deck