زواہ

ZUAH · IUB FYP 2025-2026

Hear the Truth.
Speak with Intelligence.

Urdu zero-shot voice cloning and deepfake detection. Explore the project without an account, or log in for your full workspace.

95.45%DETECTION
3 secMIN CLONE
4AI MODELS
Login Create account

About the project

Two Powerful AI Capabilities.
One Unified System.

Zuah combines zero-shot Urdu voice cloning with state-of-the-art deepfake detection. Urdu remains underrepresented in voice AI, while deepfakes spread misinformation in Urdu-speaking communities — Zuah addresses both challenges in one platform.

Built as an IUB FYP (2025–2026), the system runs on Google Colab GPU with a Flask REST API exposed through an ngrok tunnel — no local GPU required.

زواہ اردو آواز کی سچائی اور جعل سازی کا پتہ لگاتا ہے

— Zuah, FYP 2025–2026

Urdu is underserved in voice AI Deepfakes threaten audio authenticity No unified Urdu voice AI platform existed
OmniVoice

Zero-Shot Cloning

Clone any voice from 3 seconds of audio

95.45% Accuracy

Deepfake Detection

Real or fake? Know instantly.

CAPABILITIES

Everything You Need for Urdu Voice AI

From zero-shot cloning to real-time deepfake detection - Zuah handles the full pipeline.

Zero-Shot Voice Cloning

Clone any Urdu voice from just 3-10 seconds of reference audio. No speaker training required.

OmniVoice

Deepfake Detection

Multi-model fusion of Wav2Vec2, AASIST, and Conformer delivers state-of-the-art anti-spoofing performance.

95.45% Accuracy

Batch Processing

Pair up to 10 reference voices with 10 Urdu scripts and run all jobs sequentially with per-job progress tracking.

Up to 10x10

Urdu RTL Support

Full right-to-left Urdu script input using Noto Nastaliq Urdu typography with RTL-aware placeholders.

Bilingual

Real-Time Confidence Bars

Animated Real % and Fake % bars update after each analysis, giving instant visual confidence feedback.

Live Results

GPU-Accelerated Backend

Flask + PyTorch on Google Colab GPU, tunneled via ngrok for zero-infrastructure browser access.

CUDA / Colab

DEVELOPMENT PROCESS

Software Development Life Cycle

How Zuah was designed, built, tested, and deployed

Week 1-2

Requirements Gathering

  • Identified gap: no unified Urdu voice AI platform
  • Defined dual goals: cloning + detection
  • Surveyed ASVspoof 2019 benchmark for detection baseline
  • Chose zero-shot approach to avoid Urdu dataset scarcity
Week 3-4

System Design

  • Client-server architecture via ngrok tunnel
  • Selected Wav2Vec2, AASIST, Conformer for detection fusion
  • OmniVoice selected for zero-shot Urdu TTS
  • Flask REST API designed (/health, /clone, /predict)
  • Single-file HTML frontend (no build pipeline)
Week 5-10

Implementation

  • Flask backend with PyTorch model loading + CUDA inference
  • OmniVoice integrated for reference + script — WAV synthesis
  • Vanilla JS frontend: drag-drop, batch jobs, RTL Urdu UI
  • ngrok tunnel configuration for Colab deployment
  • Custom audio player and confidence bar visualizations
Week 11-13

Testing & Evaluation

  • Detection accuracy: 95.45% on evaluation set
  • Tested on ASVspoof 2019 subsets (LA and PA partitions)
  • Cross-browser testing (Chrome, Firefox, Edge, Safari)
  • Latency benchmarking: 10-30s clone / ~2s detect
  • UI testing: drag-drop, batch cloning, RTL input
Week 14

Deployment

  • Google Colab GPU as backend runtime (free tier)
  • ngrok tunnel: static subdomain for persistent URL
  • Frontend: single index.html (GitHub Pages / local)
  • Models loaded from Google Drive on Colab startup
  • End-to-end FYP demo: generate — detect workflow
0
Weeks of Development
0
Detection Accuracy
0
Lines of Code