Unmute

Inspiration

Humans are social by nature — wired, across millennia, to read faces, hold eye contact, and think on our feet in front of one another. But something has shifted. A growing number of us feel a quiet dread at the thought of an unscripted conversation, an unplanned phone call, or a room full of strangers. We think three things are driving this:

• The editable self: Texts, captions, and DMs can be drafted, deleted, and perfected before anyone sees them. Live conversation can't be undone — and that unforgiving, unedited nature has become unfamiliar, even frightening, for a generation that grew up editing every sentence before sending it.
• The scroll replaces the conversation: Social feeds are built to reward a like, a scroll, or an emoji — not a sustained back-and-forth. Every hour spent passively scrolling is an hour not spent practicing the real-time reflexes a live conversation demands.
• Comparison over connection: Curated highlight reels raise the bar for what "good enough" looks like, making an ordinary, unscripted moment with another person feel riskier than it should.

AI can draft your emails, summarize your meetings, and even hold a conversation. But it cannot sit across the table at your next interview, comfort a friend, or ask for what you deserve. Those moments still need a human voice — yours.

We built Unmute so people can use AI for exactly what it's good for: a private, judgment-free place to rehearse the conversations that matter, so they can show up fully human when it counts.

What it does

Unmute turns "I'll figure it out when I get there" into deliberate practice. It's a gamified rehearsal space for high-stakes verbal communication, built around three core pillars:

• Live AI Roleplay Calls: A fullscreen, hands-free voice call against a Gemini-powered persona (VC investor, corporate recruiter, university professor, HR director, or a fully custom persona). The AI interrupts you mid-pitch with a real objection, you respond out loud, and it fires back in character — a genuine, low-stakes rehearsal for a real high-stakes moment. Includes 4 preset scenarios (a bank risk-management interview, a PhD thesis defense, a salary negotiation, a crisis PR press conference) plus a custom-scenario generator.
• Structured Lessons: A curriculum-style lesson ("The Art of the Pause") teaches a concrete 3-step framework for handling interruptions under pressure, with explicit target benchmarks (130–150 WPM, ≤1 filler word, authoritative tone), then puts you straight into a graded practice round against the AI engine.
• Mini-Lectures: A "reading-plan"-style audio curriculum with three short audio lessons on speaking, body language, and mindset — each with a full transcript, reflection questions, and pointers to real research (cognitive reappraisal, growth mindset, posture, and nonverbal communication).

Gamification Layer: XP, day-over-day streaks, unlockable achievements (First Pitch, Zero Fillers, Streak Master), and a daily practice-credit system that gives users a reason to return and rehearse daily.

How we built it

• Frontend: Next.js 15 (App Router) + TypeScript + Tailwind CSS v4, shadcn/ui primitives, and a dark "executive" visual identity built with Tailwind gradients and glows.
• Speech-to-Text: Groq's hosted Whisper (whisper-large-v3-turbo) for rapid, reliable transcription.
• AI Persona Engine: Google Gemini 2.5 Flash, prompted dynamically per persona to stay in character, respond briefly (1–3 sentences) for natural call pacing, and track conversation history across turns.
• Real-time Voice Detection: Web Audio API's AnalyserNode, reading live microphone RMS amplitude to detect when the user starts and stops speaking — a lightweight voice-activity-detection (VAD) loop with a hard-stop safety timer.
• Voice Output: Browser-native Web Speech API (speechSynthesis) for AI spoken replies, paired with a duration-estimated fallback timer to handle native browser voice engine edge cases.
• State & Gamification: A single localStorage-backed progress store tracking XP, real day-over-day streaks, daily credits with automatic midnight resets, and achievement predicates.
• Testing: Vitest covering streak/credit math and achievement-unlock logic across fake-timer multi-day scenarios.

Challenges we ran into

• Voice-Activity Detection without calibration data: Tuning RMS thresholds for speech detection across uncalibrated microphones required building a hardened loop with an independent timeout fallback so the call never hangs.
• Web Speech API Unreliability: Browser speechSynthesis.onend events can occasionally fail to fire. We built an estimated-duration fallback timer so a stuck utterance never freezes the back-and-forth conversation loop.
• Latency & Multi-Turn Coherence: To keep latency low, we cap conversation history context to the last 6 turns and instruct Gemini to return direct conversational text instead of JSON, eliminating a full parsing round-trip.

Accomplishments we're proud of

A hands-free, real-time, two-way voice conversation with an LLM persona — with no push-to-talk button — built entirely from browser-native APIs (MediaRecorder + Web Audio + Web Speech) and zero paid voice infrastructure. Beyond the tech, we built a product grounded in real pedagogical frameworks (the Power Pause method) and nonverbal communication research.

What's next for Unmute

• Real Accounts & Persistence: Supabase authentication and PostgreSQL database integration to sync progress across devices.
• Server-Side Enforcement: Server-validated daily practice credit allocation and usage.
• Expanded Content: Additional lesson modules, mini-lecture episodes, and a community scenario-sharing hub.
• Custom Voices: Optional integrations for natural-sounding neural voice models.

Tech Stack

Next.js 15 · TypeScript · Tailwind CSS v4 · Groq Whisper (whisper-large-v3-turbo) · Google Gemini 2.5 Flash · Web Audio API · Web Speech API · Vitest

AI can already do almost anything. It cannot yet do the one thing that decides most of what happens to us — sit in the room, hold its nerve, and speak. That part is still, and will always be, uniquely human. Unmute is here to help you own it.

Built With

• gemini
• groq
• next.js
• tailwindcss
• typescript
• vitest
• webaudioapi