ClearVoice AI — AI Video Denoiser

One-click AI denoising for short-form video: make voices pop and cut background noise—free in the browser.

Traditional Chinese README: README.zh.md

Try the live demo · 简体中文 / 繁體說明

---

Why this exists

Great audio keeps people watching Reels, TikTok, and YouTube Shorts. Many creators hit the same limits:

• Environmental noise — traffic, crowds, wind
• Gear / room noise — AC hum, laptop fans
• Budget — no pro mic yet

ClearVoice AI uses DeepFilterNet3 to reduce non-speech noise while keeping voices clear—upload an .mp4 in the browser, no desktop app install.

---

Features

• One click — drag & drop .mp4, processing in the cloud
• DeepFilterNet3 — perceptual speech enhancement tuned for clarity
• Privacy-minded — stateless flow; temp uploads are cleaned up after response
• Minimal UI — Next.js-inspired dark layout, focused workflow

---

Architecture

Modern split stack: Next.js frontend (Vercel) talks directly to FastAPI backend (Docker on Hugging Face Spaces).

Browser  ──(FormData POST)──►  Next.js 14 (Vercel)
                                      │
                              NEXT_PUBLIC_API_URL
                                      │
                                      ▼
                           FastAPI (HF Spaces :7860)
                           ├─ ffmpeg: extract / resample audio (48 kHz mono)
                           ├─ DeepFilterNet3: neural denoise
                           └─ ffmpeg: mux cleaned audio → output MP4

---

Project structure

aiNoiseFilter/
├── README.md              # English (default for GitHub)
├── README.zh.md           # Traditional Chinese
├── backend/
│   ├── app.py             # POST /api/denoise
│   ├── requirements.txt   # pinned Python deps
│   └── Dockerfile         # ffmpeg + Python runtime (+ git where needed)
└── frontend/
    ├── package.json
    ├── next.config.mjs
    ├── .env.local.example
    ├── public/
    │   └── og-image.png    # Open Graph preview (replace with branded art)
    └── app/
        ├── layout.tsx     # fonts + SEO metadata
        ├── page.tsx       # main UI (client-side)
        ├── robots.ts
        ├── sitemap.ts
        └── components/

---

Local development

Prerequisites

• macOS / Linux / Windows (WSL2 recommended on Windows)
• Python 3.11+
• Node.js 18+
• FFmpeg
  • macOS: brew install ffmpeg
  • Ubuntu/Debian: sudo apt install ffmpeg

1) Backend

cd backend
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate

# Match torchaudio 2.2.x — newer releases can break DeepFilterNet imports
pip install torch==2.2.2 torchaudio==2.2.2 deepfilternet==0.5.6 fastapi uvicorn python-multipart
uvicorn app:app --reload --port 7860

2) Frontend

cd frontend
cp .env.local.example .env.local
# Edit .env.local: NEXT_PUBLIC_API_URL (backend) + NEXT_PUBLIC_SITE_URL (canonical site URL for SEO)

npm install
npm run dev

---

Deployment

Backend — Hugging Face Spaces (Docker SDK)

1. Create a Space with SDK Docker.
2. Push the contents of backend/ to the Space repo.
3. Cold starts on the free tier can take ~30–60s; first inference may download model weights (~200 MB).

Frontend — Vercel

1. Import this GitHub repository.
2. Set Root Directory to frontend.
3. Add environment variables:

   Name	Example
   NEXT_PUBLIC_API_URL	https://<your-space>.hf.space
   NEXT_PUBLIC_SITE_URL	https://clear-voice-ai.vercel.app

Rebuild after changing env vars so Open Graph URLs resolve correctly.

SEO checklist (short)

• Set NEXT_PUBLIC_SITE_URL on Vercel to your production URL.
• Replace frontend/public/og-image.png with a 1200×630 branded image.
• Submit the site URL in Google Search Console after deploy.

---

Technical notes

• torchaudio==2.2.2 — DeepFilterNet relies on APIs removed in newer torchaudio; pin versions per backend/requirements.txt.
• 48 kHz mono — ffmpeg normalizes audio before inference (DeepFilterNet’s sweet spot).
• Docker git — some DeepFilterNet logging paths assume a git repo; installing git in the image avoids noisy failures.

---

Credits & license

• Model / research: DeepFilterNet by Rikorose (see upstream license for redistribution / commercial terms).
• This repository: MIT.