Pitchfall

Free, local, open-source video transcription and subtitle generation.
Upload any video or audio file — or paste a URL — and get a timestamped transcript, synchronized playback, translation into 10 languages, and .srt subtitle export. No subscriptions, no usage limits, no data leaving your machine.

---

Screenshots

---

Why this exists

Every transcription tool online either has a paywall, a strict free tier, or sends your files to a remote server. Pitchfall runs entirely on your own machine using faster-whisper — a highly optimized local implementation of OpenAI's Whisper model. No API key required for transcription. No account. No cloud.

---

Features

• 🎙️ Local transcription — powered by faster-whisper (Whisper small model by default, CPU-friendly)
• 🔗 Video sync — click any transcript segment to jump to that exact moment in the video
• 🌍 Translation — translate the transcript into 10 languages (optional, see below)
• 📥 Export — download transcript as .txt or subtitles as .srt
• 🔗 URL support — paste any YouTube or direct video URL (powered by yt-dlp)
• 🔒 Privacy — files are processed locally and deleted immediately after transcription

---

Requirements

Dependency	Version	Notes
Python	3.10+	3.12 recommended
Node.js	18+
ffmpeg	any recent	required by faster-whisper and yt-dlp
RAM	2GB+ free	Whisper small uses ~1GB during transcription

First run: on first transcription, faster-whisper automatically downloads the Whisper small model (~245MB). This happens once and is cached locally in ~/.cache/huggingface/.

---

Installation

Option A — Docker (recommended)

git clone https://github.com/scibilo/pitchfall.git
cd pitchfall
cp .env.example .env
# Edit .env — see Configuration below
docker compose up

Open http://localhost:3000.

---

Option B — Manual setup

1. Clone

git clone https://github.com/scibilo/pitchfall.git
cd pitchfall

2. Backend

python3 -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r backend/requirements.txt

3. Frontend

cd frontend
npm install
cd ..

4. Configure

cp .env.example .env
# Open .env with any editor and fill in the values

5. Run

bash start.sh

Open http://localhost:3000. To stop: bash start.sh stop

---

Configuration

Copy .env.example to .env and edit:

# OpenRouter API key — OPTIONAL, only needed for translation.
# Transcription works fully without this key.
# Get a free key at: https://openrouter.ai/
OPENROUTER_API_KEY=

# Comma-separated list of allowed frontend origins.
ALLOWED_ORIGINS=http://localhost:3000

Variable	Required	Purpose
OPENROUTER_API_KEY	No	Enables the translation feature
ALLOWED_ORIGINS	Yes (has default)	Controls which origins can call the API

No other keys are needed. Whisper runs locally — no OpenAI key, no Google key, nothing else.

---

Translation

Translation is powered by OpenRouter using free-tier models (Gemma, Llama, Mistral, DeepSeek, Qwen — tried in order with automatic fallback).

Important things to know:

• Free models have rate limits. A 503 - All translation models are temporarily unavailable error means the free tier is momentarily saturated. Wait 10–30 seconds and retry — this is not a bug.
• Free models occasionally go offline. Pitchfall tries 5 different models before giving up, so partial outages are handled automatically.
• Translation is optional. Transcription, video sync, and .srt export all work without any API key.
• Want reliable translation? Upgrade to a paid plan on OpenRouter and change the model name in backend/services/translation_service.py. The code structure stays identical. At OpenRouter's current pricing, a few hundred transcriptions cost less than $1.

---

Usage

1. Open http://localhost:3000
2. Upload a file — drag & drop or browse (MP4, MP3, MOV, WAV, and most common formats)
3. Or paste a URL — YouTube, Vimeo, or any direct media URL supported by yt-dlp
4. Wait for transcription — a progress bar shows the current stage and latest recognized segment
5. Click any segment in the transcript panel to jump to that moment in the video
6. Translate (optional) — select a language and click Translate
7. Export — copy to clipboard, download as .txt, or download subtitles as .srt

---

Project structure

pitchfall/
├── backend/                      # FastAPI + faster-whisper
│   ├── main.py                   # API endpoints, startup/shutdown cleanup
│   ├── requirements.txt
│   └── services/
│       ├── transcription_service.py   # Whisper streaming transcription
│       ├── translation_service.py     # OpenRouter with 5-model fallback
│       └── ytdlp_service.py           # URL media download
├── frontend/                     # Next.js 16 + Tailwind CSS 4
│   ├── postcss.config.mjs        # Required for Tailwind CSS 4
│   ├── next.config.ts
│   ├── package.json
│   └── src/
│       ├── app/
│       │   ├── layout.tsx
│       │   ├── page.tsx          # Main page, state management, reset logic
│       │   └── globals.css
│       └── components/
│           ├── TranscriptViewer.tsx   # Video player + transcript sync
│           ├── UploadForm.tsx
│           ├── UrlForm.tsx
│           └── ProgressBar.tsx
├── docker-compose.yml
├── start.sh                      # Dev launcher (no Docker required)
├── cleanup.sh                    # Clears temp files and build caches
└── .env.example

---

Changing the Whisper model

In backend/services/transcription_service.py, change:

model_size = "small"

Model	Download size	RAM	Speed (CPU)	Accuracy
tiny	~75MB	~390MB	fastest	lower
base	~145MB	~500MB	fast	decent
small	~245MB	~1GB	moderate	good (default)
medium	~770MB	~2GB	slow	very good
large-v3	~1.5GB	~3GB	very slow	best

For everyday use on a laptop, small is the right tradeoff. For professional subtitling, medium or large-v3 deliver significantly better accuracy (GPU strongly recommended for those sizes).

---

Troubleshooting

Page loads but has no styling (plain HTML)
The postcss.config.mjs file is missing from frontend/. Add it:

const config = { plugins: { "@tailwindcss/postcss": {} } };
export default config;

Then delete frontend/.next/ and restart.

Translation returns 503
Free OpenRouter models are rate-limited or temporarily offline. Wait 10–30 seconds and retry. Check status.openrouter.ai for ongoing outages.

Transcription hangs at 0%
The Whisper model is downloading for the first time (~245MB). Check the terminal for download progress — it only happens once.

ffmpeg: command not found
sudo apt install ffmpeg on Ubuntu/Debian, or brew install ffmpeg on macOS.

CORS error in browser console
Set ALLOWED_ORIGINS in .env to match the exact URL your frontend runs on.

---

Security

Pitchfall is designed for local, single-user use on a trusted machine. It is not hardened for public deployment. Before exposing it on a network:

• No authentication. Any client reaching the backend port can upload files and consume your OpenRouter credits.
• No file size limits. A large upload can fill your disk.
• CORS defaults to http://localhost:3000. Do not widen it without also adding authentication.
• .env must never be committed. Run git status before every push and confirm .env is not in the staged files. If a key is leaked, rotate it on the OpenRouter dashboard immediately.
• For anything beyond local use, deploy behind a reverse proxy (nginx/Caddy) with HTTPS, HTTP basic auth, and rate limiting.

---

Cleanup

bash cleanup.sh

Removes leftover temp files, Python __pycache__, and Next.js build caches.

---

Contributing

Pull requests are welcome. For major changes, open an issue first to discuss what you'd like to change.

---

License

MIT — see LICENSE.