Hushmark

Describe a sound. We'll erase it.

Text-prompted audio source separation that removes exactly what you describe—barking dogs, coughs, sirens, keyboard clicks—from your audio and video files.

How It Works

1. Upload your audio (WAV, MP3, M4A) or video (MP4, MOV)
2. Describe the unwanted sound: "dog barking", "keyboard clicks", "sirens"
3. Download your cleaned file + the isolated removed sound

Optional: Mark 1-3 example regions in the waveform for surgical precision.

Tech Stack

Layer	Technology
Frontend	Next.js 15, TypeScript, Tailwind, shadcn/ui
Backend	Convex (realtime database + actions)
GPU Inference	Modal (serverless), AudioSep model
Storage	Cloudflare R2
Deployment	Vercel

Development

# Install dependencies
pnpm install

# Start Convex dev server (terminal 1)
pnpm convex:dev

# Start Next.js dev server (terminal 2)
pnpm dev

# Deploy Modal worker
pnpm worker:deploy

Project Structure

apps/
  web/           # Next.js frontend + Convex backend
  worker/        # Modal GPU worker (Python)
packages/
  config/        # Shared TypeScript/ESLint configs

Environment Variables

Convex Dashboard

CLOUDFLARE_ACCOUNT_ID=xxx
R2_ACCESS_KEY_ID=xxx
R2_SECRET_ACCESS_KEY=xxx
R2_BUCKET=hushmark
MODAL_WORKER_URL=https://your-modal-endpoint
WORKER_CALLBACK_SECRET=random-secret

Modal Secret (hushmark-r2)

R2_ENDPOINT=https://xxx.r2.cloudflarestorage.com
R2_ACCESS_KEY_ID=xxx
R2_SECRET_ACCESS_KEY=xxx

Architecture

Browser → Presigned PUT → R2 (input)
       ↓
    Convex (job created)
       ↓
    Modal Worker (GPU inference)
       ↓
    R2 (outputs) → Presigned GET → Browser
       ↓
    Convex callback (job complete)

See DESIGN.md for comprehensive architecture documentation.

License

Private - All rights reserved.