CVR — Cockpit Voice Recorder

AI-powered video subtitle generator. Transcribe any video with Whisper and translate subtitles using OpenAI or Gemini.

Available as a CLI tool and a desktop app (Windows / macOS).

video → ffmpeg → Whisper (GPU) → LLM translate → .srt

Features

• GPU-accelerated transcription — faster-whisper with CUDA support (auto-detects GPU)
• Multiple Whisper models — small, medium, large-v3 (choose speed vs accuracy)
• Multi-provider translation — OpenAI (GPT-4o-mini) or Google Gemini (2.5 Flash)
• Bilingual subtitles — original + translated text in one .srt file
• 7 languages — English, Chinese, Japanese, Korean, Spanish, French, German
• Desktop app — drag & drop UI built with Tauri + React, aviation-inspired dark theme

Desktop App

The desktop app provides a visual interface with real-time processing progress.

Requirements

• Python 3.10+ with cvr CLI installed
• ffmpeg in PATH
• NVIDIA GPU recommended (CUDA) — falls back to CPU if unavailable

Install & Run

# Install the CLI backend
pip install cvr

# Run the desktop app (dev mode)
cd desktop
npm install
npm run tauri dev

Build Installer

cd desktop
npm run tauri build
# Output: desktop/src-tauri/target/release/bundle/

CLI

Install

pip install cvr

Windows note: If cvr is not recognized after install, use python -m cvr instead.

Usage

macOS / Linux:

# Auto-detect language, translate to Chinese (bilingual)
cvr movie.mp4

# Translate to Japanese
cvr movie.mp4 -t ja

# Use Gemini for translation
cvr movie.mp4 --provider gemini

# English source, Chinese translation only
cvr movie.mp4 -l en -t zh --mode translated

# Transcribe only, no translation
cvr movie.mp4 --no-translate

# Smaller model on CPU
cvr movie.mp4 -m small --device cpu

Windows (PowerShell):

# Use python -m cvr if cvr.exe is not in PATH
python -m cvr movie.mp4

# Set API key and translate
$env:OPENAI_API_KEY = "sk-xxx"
python -m cvr movie.mp4

# Or use Gemini
$env:GEMINI_API_KEY = "AIza-xxx"
python -m cvr movie.mp4 --provider gemini

# Or pass API key directly
python -m cvr movie.mp4 --provider gemini --api-key "AIza-xxx"

# Network paths (UNC) — use forward slashes
python -m cvr "//server/share/movie.mp4"

Options

Option	Default	Description
-o, --output	{video}.{lang}.srt	Output file path
-l, --language	auto-detect	Source language code
-t, --target	zh	Target translation language
--provider	openai	openai or gemini
--llm	auto	Model name (default: gpt-4o-mini / gemini-2.5-flash)
--api-key	env var	API key (alternative to environment variable)
-m, --model	large-v3	Whisper model: small, medium, large-v3
--mode	bilingual	bilingual, translated, or original
--device	auto	cuda, cpu, or auto
--no-translate		Transcribe only
--json-progress		Output JSON lines (for GUI integration)

Whisper Models

Model	Size	VRAM	Speed	Best For
small	461MB	~1GB	Fastest	Quick drafts
medium	1.5GB	~2GB	Balanced	Most use cases
large-v3	2.9GB	~6GB	Slowest	Maximum accuracy, rare languages

Environment Variables

# OpenAI (for translation)
export OPENAI_API_KEY=sk-xxx

# Gemini (for translation)
export GEMINI_API_KEY=AIza-xxx

Output Example

Bilingual mode (default — original + translated):

1
00:00:02,800 --> 00:00:06,000
[Original language line]
[Translated language line]

2
00:00:06,000 --> 00:00:11,700
[Original language line]
[Translated language line]

Translated only (--mode translated):

1
00:00:02,800 --> 00:00:06,000
[Translated language line only]

How It Works

1. ffmpeg extracts audio from video (16kHz mono WAV)
2. faster-whisper transcribes with GPU acceleration and timestamps
3. LLM (OpenAI or Gemini) translates in batches of 30 segments for context coherence
4. Generates standard .srt subtitle file

Tech Stack

• CLI: Python, faster-whisper, ffmpeg, OpenAI SDK, Google GenAI SDK
• Desktop: Tauri 2 (Rust) + React + TypeScript + Tailwind CSS
• GPU: CUDA via PyTorch + CTranslate2

Project Structure

cvr/
├── src/cvr/           # Python CLI
│   ├── cli.py         # CLI entry point
│   ├── audio.py       # ffmpeg audio extraction
│   ├── transcribe.py  # Whisper transcription
│   ├── translate.py   # LLM translation (OpenAI / Gemini)
│   └── srt.py         # SRT file generation
├── desktop/           # Tauri desktop app
│   ├── src/           # React frontend
│   └── src-tauri/     # Rust backend
├── pyproject.toml
├── LICENSE            # MIT
└── README.md

Disclaimer

This software is provided as a general-purpose transcription and translation tool. Users are solely responsible for ensuring their use complies with all applicable laws and regulations, including but not limited to:

• Copyright law: Generating subtitles for copyrighted content may require authorization from the rights holder. This tool does not grant any rights to the underlying media.
• Fair use / fair dealing: Whether subtitle generation constitutes fair use depends on your jurisdiction and specific circumstances. Consult legal counsel if uncertain.
• Distribution: Subtitles derived from copyrighted works may themselves be subject to copyright restrictions. Redistribution of generated subtitle files is at your own risk.
• Accuracy: Machine-generated transcriptions and translations may contain errors. The authors make no guarantees regarding the accuracy, completeness, or fitness for any purpose of the generated output.

The authors and contributors of CVR accept no liability for any content generated using this software, any copyright infringement arising from its use, or any damages resulting from reliance on its output. Use at your own risk.

By using this software, you acknowledge that you have the legal right to process the media files you provide.

License

MIT