Nani's Recipe Book: Turning Voice Memos into a Family Recipe Book with Gemma 4

Bhavya Gupta — Sun, 04 Oct 2026 13:44:21 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

🍲 What I Built

Every family has that one person whose food tastes like home. 👵❤️ For us, it's my [Nani]. She knows dozens of recipes by heart, but she has never written a single one down. No measuring cups, no written steps, just "thoda sa namak, thoda sa adrak, aur bas pakne do."

I was scared that one day those recipes would be gone. So I built Nani's Recipe Book 🥘

Nani doesn't need to type anything. She just talks, the way she always does, and the app:

🎙️ Listens to her voice memo (in Hindi or English)
📝 Writes it down so the family can fix any mistakes
🍛 Turns it into a clean, colourful recipe card with ingredients and numbered steps

My real story: the dish which is my favourite is her daal makhni recipe that makes the quality of her food is super delicious. 💛

📸 Demo

Before: a spoken Hindi memo. After: a recipe with ingredients and numbered steps. ✨

🎬 Video

▶️ Watch the demo video

https://screenrec.com/share/pU4sBVQm0t

💻 Code

👉 [https://github.com/Bhavy12-cell/Granny_Recipes.git]

🛠️ How I Built It

🎧 Speech to text: Whisper large-v3 (open-weight) via Groq
🍳 Recipe formatting: Gemma 4 31B (open-weight) via Google AI Studio API
🎨 UI: Streamlit with custom CSS
🔁 Flow: voice memo → editable transcript → Gemma → JSON recipe → recipe book

Recipes are family memories, so I did not want the AI to "improve" them. The prompt tells Gemma to:

use only what was said and never invent an ingredient
keep vague amounts like "thoda sa" exactly as spoken
never translate away from Nani's own language

The transcript is editable before formatting, so the family always has the final word. When the API returned a 500 error during testing, I added automatic retries and a fallback to a second Gemma model.

🌍 Why Does Open Innovation Matter?

Both models in this project are open-weight, and that shaped how I built it:

🔓 Freedom to swap and tune. I could try different Gemma sizes and tune the prompt for Hindi and family-style recipes. With a closed model I would have had to accept whatever it did.
🏠 A path to full privacy. Family voice recordings are personal. Because the models are open, the same pipeline can run fully on a laptop (for example with Ollama and faster-whisper) so audio never leaves the house.
💸 Low cost. The whole project runs on free tiers, which matters when you are a student on a limited data plan.

An honest tradeoff: I first tried the local route with Ollama, but my mobile data ran out while downloading the model. So this version uses hosted APIs, which means the audio and transcript do leave my laptop. Open models are what made the switch easy: I kept the same open models and only changed where they run. A fully local version is my next step. 🔒

🔮 What's Next

🔒 Fully local version for privacy
📖 Printable recipe book (PDF) to gift to the family
🔊 A "read aloud" button so Nani can listen to her own recipes

🏆 Prize Categories

Best Use of Gemma 💎

Made with love, for the best cook I know. 🧡

New to community

Bhavya Gupta — Sat, 03 Oct 2026 10:43:01 +0000

Hello Everyone, Nice to meet you

DEV Community: Bhavya Gupta