Tablespeak

Translate a restaurant menu into any Indian language, and keep every dish spelled the way you serve it.

A generic translator guesses dish names. It turns ઊંધિયું into "mixed vegetable curry", spells khaman three different ways on one menu, and invents ingredients. For a restaurant that is proud of its food, that is the wrong menu to hand a tourist.

Tablespeak gives the model the restaurant's own glossary, a dish dictionary the owner controls. Dishes in the glossary keep the exact spelling and description the owner chose, in every language. Everything else is translated normally.

ઊંધિયું / undhiyu | Undhiyu | Surat's winter specialty: mixed vegetables slow-cooked with green masala
└─ as written on the menu   └─ preferred spelling   └─ description the model must use

Think of it as handing a freelance translator the restaurant's cookbook before they start.

flowchart LR
    A["Menu text<br/>Gujarati, Hindi or English"] --> S["Tablespeak server"]
    G["Restaurant glossary"] --> S
    S -->|"prompt + glossary"| M["Open-weight model<br/>Qwen, Llama or Gemma"]
    M -->|"JSON"| S
    S -->|"script check, retry once if needed"| C["Printable menu card"]

The API key lives only on the server, so visitors never see it.

You need Node 18+ and one API key from a provider that hosts open-weight models (OpenRouter works well).

git clone https://github.com/Jainil26/tablespeak.git
cd tablespeak
cp .env.example .env      # Windows: copy .env.example .env
# open .env and paste your key into LLM_API_KEY
npm start

Open http://localhost:3000, click Use a sample menu, then Translate menu. There is no npm install step.

1. Import the repo in Vercel. No framework preset, no build command.
2. Add LLM_API_KEY and LLM_BASE_URL (https://openrouter.ai/api/v1).
3. Deploy. Set a spending limit on your key before you share the link.

• Swap in one line. Change the model list in lib/config.js, or set LLM_MODELS, and compare Qwen, Llama and Gemma on the same menu.
• No lock-in. Any OpenAI-compatible endpoint works. Change LLM_BASE_URL and nothing else.
• Private when it matters. Point the same code at a local Ollama server on a stronger machine, and menu data never leaves the building.
• Cheap. A menu is a few thousand tokens, so a translation costs very little, and open models are often the cheapest per token.
• Yours to change. The glossary and prompt are plain text, and open weights can be fine-tuned later for a specific cuisine.

Model ids change over time. Check your provider's model list if one returns an error.

public/index.html     the whole front end (HTML, CSS, JS)
public/glossary.txt   starter glossary of Gujarati dishes
lib/translate.js      prompt, model call, JSON parsing, script check and retry
lib/config.js         languages and model list
lib/handlers.js       request handling and a simple rate limit
api/                  Vercel serverless functions
server.js             local dev server, no dependencies

• Output is AI-generated. Check ingredients and allergens before printing.
• Open models are stronger in Hindi, Gujarati, Marathi and Bengali than in Malayalam, Kannada or Telugu. Compare models for your language.
• The rate limit is in memory and best effort. Cap your API spend if the site is public.

Contributions are welcome, and the repo is open for Hacktoberfest. Good first issues:

• Add a glossary starter pack for another cuisine (Punjabi, South Indian, Bengali, Rajasthani)
• Add more languages, such as Odia, Assamese or Urdu
• Upload a menu as a photo or PDF and extract the text first
• Add automated tests for lib/translate.js
• Import and export the glossary as CSV
• Add a dark-text print stylesheet for black-and-white printers

Open an issue to discuss your idea, then send a pull request.

MIT. Built by @Jainil26 for the Hacktoberfest 2026 open-source AI challenge, Side Hustle theme.