subai

SubtitleAI

Seamlessly clean up and convert Zoom's VTT subtitles to Markdown with the power of OpenAI.

This command-line tool converts VTT files (used by Zoom for subtitles) into a human-readable format, utilizing the power of OpenAI's GPT model. It makes the transcript much more coherent and ready for subsequent tasks (e.g., summarization).

Prerequisites

Before using this tool, make sure you have the following:

• Node.js installed on your system.

Usage

To process a VTT file, use the following command:

export OPENAI_API_KEY=sk-***
npx subai input.vtt output.md

Make sure to replace sk-*** with your OpenAI API key,input.vtt with the path to your VTT file, and output.md with the desired path for the generated Markdown file. The API key will be saved temporarily for each shell session, so you only need to re-provide the API for each new shell session (e.g., you opened a new terminal window).

Environment Variables

SubAI can be customized with the following environment variables:

• OPENAI_API_MODEL (defaults to gpt-3.5-turbo): The code name of the GPT model to use. You can find available models in the OpenAI API documentation.
• OPENAI_API_MAX_TOKENS (defaults to 2048): The maximum number of tokens per chunk. You can set this variable to limit the size of the generated Markdown chunks. If not provided, the default value will be used.
• SUBAI_PROMPT: Custom prompt for AI. The default prompt is designed to make minimal changes to the original transcript. See openAIProcessor.js for the default prompt.

To set an environment variable, run:

export VARIABLE=VALUE

Additional Commands

In addition to the main vtt2md command, this tool provides the following commands:

• npx --package=subai vtt-clean: Cleans a VTT file by removing redundant data and timestamps to save tokens. This will be used by ai-process below.
• npx --package=subai ai-process: Processes the cleaned file with OpenAI, generating a Markdown file.

npx subai essentially chains these two commands together.

Contributing

Contributions are welcome! If you find any issues or have suggestions for improvements, please open an issue on the GitHub repository.

License

This tool is licensed under the ISC License. See the LICENSE file for more information.

Author

This tool was created by Zenan (Alan) Chen. You can find more of his work on his homepage.