UrduNexus: AI Transliteration Engine

UrduNexus is a high-speed, AI-powered system that seamlessly converts Roman Urdu text into native Urdu script. Built using a fine-tuned ByT5 model, it features a self-contained architecture hosted on Google Colab, completely bypassing local hardware constraints and network blocks.

Key Features

• Real-Time Transliteration: Converts Roman Urdu to native script instantly using a customized google/byt5-small model.
• Smart Chunking: Automatically handles long paragraphs without hitting token limits.
• Voice Integration: Built-in Text-to-Speech (TTS) using Microsoft edge-tts (ur-PK-UzmaNeural voice).
• Self-Contained Architecture: The Python backend dynamically generates the HTML/CSS/JS frontend on the fly—no manual file uploads required.
• Unblockable Tunneling: Utilizes bore.pub to expose the local Colab FastAPI server to the public web securely.
• Glassmorphism UI: A beautiful, responsive, and modern web interface.

Tech Stack

• Machine Learning: PyTorch, Hugging Face Transformers (Maira26/UrduNexus-ByT5)
• Backend: Python, FastAPI, Uvicorn
• Frontend: HTML5, CSS3, Vanilla JavaScript
• Deployment: Google Colab T4 GPU, Bore.pub Tunneling

How to Run (1-Click Setup)

Because this project is designed to run on free cloud GPUs, installation is entirely automated.

1. Open the UrduNexus_Demo.ipynb notebook in Google Colab.
2. Go to Runtime > Change runtime type and ensure T4 GPU is selected.
3. Click Run All (or press Ctrl + F9).
4. Wait 1-2 minutes for the system to install dependencies, load the model into VRAM, and spin up the tunnel.
5. Click the public bore.pub link printed at the bottom of the cell to access the live web app!

System Architecture

1. Frontend: User types Roman Urdu into the browser UI. JavaScript sends a POST request to the backend.
2. Backend Engine: FastAPI receives the text, chunks it into <180 character segments, and feeds it to the ByT5 model via PyTorch.
3. Audio Generation: The native Urdu output is hashed and converted to an .mp3 using edge-tts.
4. Response: The system returns the Urdu text and the relative path to the audio file, which the browser immediately displays and loads into the hidden audio player.

Demonstration

---