shakeeb08/youtube-to-roman-urdu

🎧 YouTube to Roman Urdu Converter

An AI-powered web application that converts Urdu and English YouTube videos into Roman Urdu using Speech Recognition and Large Language Models.

🔗 Live App: https://huggingface.co/spaces/shakeeb08/youtube-to-roman-urdu

🚀 Project Overview

This application automates the following pipeline:

YouTube URL
↓
Audio Extraction (yt-dlp)
↓
Speech-to-Text (Whisper ASR)
↓
LLM Processing (Gemini)
↓
Roman Urdu Transcript Output

The system supports:

• —Urdu audio → Roman Urdu
• —English audio → Roman Urdu

🧠 Core Features

• —✅ Automatic language detection (Urdu / English)
• —✅ Chunk-based LLM processing to avoid token overflow
• —✅ Clean, interactive Streamlit interface
• —✅ Modular architecture (downloader, transcriber, romanizer)
• —✅ Secure API key management using environment variables
• —✅ Docker-based deployment support

🏗 Architecture Design

The application follows a modular pipeline architecture:

1️⃣ Downloader Module

• —Uses yt-dlp to extract audio from YouTube videos.

2️⃣ Transcriber Module

• —Uses faster-whisper for efficient speech recognition.
• —Detects source language automatically.

3️⃣ Romanizer Module

• —Uses Google Gemini LLM.
• —Converts:
• —Urdu script → Roman Urdu
• —English → Roman Urdu
• —Implements chunking to handle long transcripts safely.

⚙️ Tech Stack

• —Python
• —Streamlit
• —faster-whisper
• —yt-dlp
• —Google Gemini API
• —Docker (for cloud deployment)

📸 Application Screenshots

🏠 Home Interface

📝 Sample Output (Roman Urdu Transcript)

📦 Local Installation Guide

1️⃣ Clone Repository

git clone https://github.com/Sadat-Shakeeb/multilingual-youtube-transcriber.git
cd multilingual-youtube-transcriber

2️⃣ Create Virtual Environment

python -m venv yt_env
yt_env\Scripts\activate   # Windows

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Add Gemini API Key

Create a .env file:

GEMINI_API_KEY=your_api_key_here

5️⃣ Run Application

streamlit run app.py

🌍 Deployment Notes

The app has been deployed on Hugging Face Spaces.

While deploying, the following real-world constraints were encountered:

• —File watcher (inotify) limits on cloud infrastructure
• —JavaScript runtime requirements for yt-dlp
• —Network restrictions on some free-tier environments

These were addressed via:

• —Streamlit configuration tuning
• —Deployment strategy adjustments
• —Modular system redesign

🎯 Key Engineering Learnings

• —Handling cloud deployment limitations
• —Managing environment secrets securely
• —Designing chunk-based LLM pipelines
• —Separating data ingestion from ML processing
• —Debugging container health checks
• —Optimizing Whisper performance on CPU

📌 Resume-Ready Project Description

Multilingual AI web application that converts Urdu and English YouTube videos into Roman Urdu using Whisper ASR and Gemini LLM. Designed with chunk-based LLM processing to handle long transcripts and deployed using Streamlit with secure API key management.

🔮 Future Improvements

• —Direct audio file upload support
• —Support for additional languages
• —Migration to updated Gemini SDK
• —Improved transcription accuracy tuning
• —Background job processing for long videos

📄 License

MIT License