VideoScript 1.0.0

VideoScript

A powerful .NET Global Tool for transcribing video and audio files to text using Apple Silicon GPU acceleration.

Features

• Automatic Video/Audio Detection: Automatically detects input file type and extracts audio from video files
• GPU Acceleration: Uses MLX-Audio framework with Apple Silicon GPU for fast transcription
• Qwen3-ASR Model: High-quality speech recognition supporting multiple languages
• Smart Paragraph Formatting: Automatically formats text with line breaks at natural pause points
• First-Run Setup: Automatically installs FFmpeg, creates Python virtual environment, and downloads the model
• ModelScope Support: Faster model downloads in China region

Installation

dotnet tool install -g VideoScript

Usage

Basic Usage

# Transcribe a video file
videoscript input.mp4

# Transcribe an audio file
videoscript input.mp3

# Specify output file
videoscript input.mp4 -o output.txt

# Specify language (zh, en, etc.)
videoscript input.mp4 --language zh

Advanced Options

# Use a different model
videoscript input.mp4 -m mlx-community/Qwen3-ASR-0.6B-4bit

# Disable text formatting
videoscript input.mp4 --no-format

# Verbose output
videoscript input.mp4 --verbose

# Force re-initialization
videoscript input.mp4 --force-init

Utility Commands

# Show help
videoscript --help

# Show version
videoscript --version

# Check environment status
videoscript status

# List available models
videoscript models

Requirements

• Operating System: macOS with Apple Silicon (M1/M2/M3/M4)
• Runtime: .NET 8.0
• Python: Python 3.10+ (automatically configured)
• Package Manager: Homebrew (for FFmpeg auto-installation)

First Run

On first run, VideoScript will automatically:

1. Install FFmpeg via Homebrew (if not already installed)
2. Create a Python virtual environment in ~/.videoscript/venv
3. Install required Python packages (mlx-audio, huggingface_hub, modelscope)
4. Download the Qwen3-ASR-1.7B-8bit model (~2.5GB) to ~/.videoscript/models/

This initialization process may take several minutes depending on your internet connection.

Output Format

The transcription output is automatically formatted with:

• Line breaks at sentence boundaries (based on Chinese and English punctuation)
• Paragraphs grouped by ~150 characters or 2+ sentences

Example output:

欢迎回到 Google 与 Kaggle 联合推出的 5-day AI Agents Intensive 白皮书解读。
前面三天我们已经打好了基础。
Day 1，我们定义了 agent 的架构基石。Day 2，通过 MCP 协议让 agent 学会了行动。

Available Models

• mlx-community/Qwen3-ASR-1.7B-8bit (default) - Best quality
• mlx-community/Qwen3-ASR-1.7B-4bit - Smaller memory footprint
• mlx-community/Qwen3-ASR-0.6B-4bit - Fastest, lowest memory
• mlx-community/whisper-small-asr-6bit - Alternative model

Performance

On Apple Silicon M-series chips:

• Audio extraction: ~2 seconds for a 10-minute video
• Transcription: ~60 seconds for a 10-minute audio (using 1.7B model)

License

MIT License

Support

For issues and feature requests, please visit: https://github.com/videoscript/videoscript/issues