I Turned My Grandmother's Voice Memos Into a Family Recipe Book, Fully Offline

Many family recipes are only kept in someone's mind. Asking about an ingredient's amount might result in vague answers like "a handful" or "until it tastes right." Once the cook is gone, the recipe disappears along with them. For a challenge, I developed Recipe Keeper for a family member who cooks from memory and rarely writes their recipes.

This tool records someone explaining a dish and converts that recording into a written recipe. All recipes are compiled into one book, and everything runs on my laptop without any internet or paid APIs. The target audience is a family member whose cooking skills I wouldn't want to lose. They often explain dishes in multiple languages, estimate quantities, and jump between steps in a recipe.

A regular recipe app wouldn't work for them, but a voice memo would. Here's how it operates: Record the dish by having the person explain it out loud and save the recording in a folder. Transcribe the audio to text using faster-whisper, an open-source version of Whisper, locally. Structure the rambling transcript into JSON format with a title, ingredients, steps, and personal tips using a local open-weight model (Gemma, served through Ollama).

Publish all the recipes into one Markdown recipe book using a small Python script. The core of the script is about 50 lines of Python. Some challenges include vague measurements, mixed languages, and invented details. To address these issues, the prompt and pipeline were designed to keep vague amounts vague, prevent the model from inventing numbers or details, and handle language switching.

Open source was crucial for privacy and to ensure the tool works offline without any cost. The code is available on GitHub for anyone to use and modify.