---
title: Arabic Nativity Classifier
emoji: 🎙️
colorFrom: indigo
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
license: mit
---

# Arabic Native vs Non-Native Speaker Detection

**Wav2Vec2 XLS-R Prosody-Based Classification System**

This application automatically classifies Arabic speakers as **Native** or **Non-Native** using advanced prosodic analysis with Meta's XLS-R multilingual speech model.

## ✨ Features

- 🎤 **Multiple Input Modes**
  - Upload CSV with batch audio URLs
  - Custom CSV column mapping
  - Direct audio file upload (MP3, WAV, etc.)
  - Single URL classification

- 🧠 **Advanced ML Pipeline**
  - Wav2Vec2 XLS-R (1.2B params) prosody layer extraction
  - 10-second audio chunking with overlap
  - Pitch, rhythm, and temporal variance features
  - MLP classifier with confidence scoring
  - Weighted chunk aggregation

- 📊 **Comprehensive Results**
  - Real-time processing progress
  - Per-sample predictions with confidence
  - Aggregate metrics (accuracy, F1, precision, recall)
  - Confusion matrix visualization
  - Chunk-level detail breakdown

- 🌍 **Supported Arabic Dialects**
  - Saudi Arabian (Arabic_SA)
  - UAE (Arabic_AE)
  - Qatari (Arabic_QA)
  - Kuwaiti (Arabic_KW)

## 🔬 Technical Details

**Model Architecture:**
- Encoder: `facebook/wav2vec2-large-xlsr-53` (XLS-R 300M)
- Prosody layers: 4, 6, 8, 10 (early-to-mid)
- Features: mean + std + F0 (pitch) → 2051-d embedding
- Classifier: 3-layer MLP (384→256→128→2)
- Threshold: 0.4 (optimized on validation)

**Pipeline:**
```
Audio → Librosa 16kHz → 10s chunks → XLS-R layers → 
Prosody embedding → L2 norm → MLP → Confidence-weighted vote
```

## 🚀 Usage

1. **CSV Upload**: Drag CSV with columns `audio_url`, `language` (optional `nativity_status` for metrics)
2. **Custom Mapping**: Upload any CSV and map columns via dropdowns
3. **Audio File**: Drop MP3/WAV and select language
4. **URL**: Paste direct audio link

The system processes in real-time with visual progress updates.

## 📈 Performance

Validation metrics on Arabic dialect corpus:
- **Accuracy**: 85.2%
- **F1 Macro**: 84.7%
- **Native F1**: 88.3%
- **Non-Native F1**: 81.1%

## 🏗️ Architecture

Built with:
- **Backend**: Flask + Gunicorn
- **ML**: PyTorch + Transformers + Librosa
- **Frontend**: Vanilla JS with Server-Sent Events
- **Deployment**: Docker multi-stage build

## 🎯 Use Cases

- Language proficiency assessment
- Call center accent analysis
- Second language acquisition research
- Speech therapy progress tracking
- Dialect variation studies

## 📝 Credits

Developed for **Hackenza 2026** — Arabic Speech Processing Challenge

Model: Meta AI XLS-R (Conneau et al., 2021)

---

**Note**: First prediction may take ~30s as models load into memory. Subsequent requests are instant.