vikrant-vikram/INDICA

# INDICA: An Audio Indic-Language Telecom Fraud Analysis Benchmark

<p align="center">
<b>Multilingual | Audio + Text | Benchmark for Fraud Detection</b>
</p>

<p align="center">
<img src="https://img.shields.io/badge/Dataset-189K%2B%20Samples-blue">
<img src="https://img.shields.io/badge/Languages-10-green">
<img src="https://img.shields.io/badge/Tasks-3-orange">
<img src="https://img.shields.io/badge/Modalities-Audio%20%2B%20Text-purple">
</p>

---

## Overview

INDICA is a comprehensive benchmark for telecom fraud call analysis in Indic languages.
It is built on the IndiF dataset, the first large-scale multilingual dataset for fraud detection in telecom conversations.

This benchmark enables research in:
- Scenario Classification
- Fraud Call Detection
- Fraud-Type Classification

INDICA addresses key challenges:
- Lack of multilingual datasets
- Limited reproducibility
- Poor cross-lingual generalization

---

## Key Contributions

- **IndiF Dataset**
- 189,420 samples across 10 Indic languages
- Audio + text modalities
- Fine-grained annotations

- **INDICA Benchmark**
- Evaluation across speech, text, and multimodal models
- Standardized experimental setup

- **Multilingual Setup**
- Enables robust cross-lingual evaluation

---

## Dataset: IndiF

### Languages
Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Odia, Tamil, Telugu

### Dataset Statistics

| Property | Value |
|---------|------|
| Total Samples | 189,420 |
| Languages | 10 |
| Samples per Language | 18,942 |
| Modalities | Audio + Text |
| Tasks | 3 |

### Annotations

- Scenario Classification (7 categories)
- Fraud vs Non-Fraud (Binary)
- Fraud-Type Classification (7 types)

---

## Tasks

### Scenario Classification
- Customer Service
- Delivery
- Ride-hailing
- Retail Transactions
- Appointment Scheduling
- Food Ordering
- Traffic Information

### Fraud Detection

```
Fraud vs Non-Fraud
```

### Fraud-Type Classification

- Banking Fraud
- Customer Service Impersonation
- Investment Scam
- Phishing
- Lottery Scam
- Kidnapping
- Identity Theft

---

## Dataset Files

- Gujarati_audio.tar.gz — Gujarati audio
- Hindi_audio.tar.gz — Hindi audio
- Tamil_audio.tar.gz — Tamil audio
- Text_samples.tar.gz — Text data
- .......

---

## How to Use

### Download

```bash
hf download vikrant-vikram/INDICA Gujarati_audio.tar.gz
hf download vikrant-vikram/INDICA Hindi_audio.tar.gz
hf download vikrant-vikram/INDICA Tamil_audio.tar.gz
hf download vikrant-vikram/INDICA Text_samples.tar.gz
......
```

### Extract

```bash
tar -xzf Gujarati_audio.tar.gz
tar -xzf Hindi_audio.tar.gz
tar -xzf Tamil_audio.tar.gz
tar -xzf Text_samples.tar.gz
....
```

### Structure

```
Audio_samples/
├── Gujarati/
├── Hindi/
├── Tamil/
....

Text_samples/
├── Gujarati/
├── Hindi/
├── Tamil/
....
```

---

## Authors

- Nitin Choudhury
- Samyuktha Chilaka
- Bikrant Bikram Pratap Maurya
- Arun Balaji Buduru

---

## Citation

```bibtex
@inproceedings{indica2026,
title={INDICA: An Audio Indic-Language Telecom Fraud Analysis Benchmark},
author={Choudhury, Nitin and Chilaka, Samyuktha and Maurya, Bikrant and Buduru, Arun},
year={2026}
}
```