JesseHuang922/VoxSentinel_Synthetic_Detection_Dataset

---
license: cc-by-nc-4.0
task_categories:
- audio-classification
tags:
- deepfake-detection
- wav2vec2
- speech-security
language:
- en
- ja
size_categories:
- 10K<n<100K
---

# Wav2Vec2-Deepfake-Sentinel-Base-Dataset

## 🛡️ Project Overview
This dataset is a core component of the **Wav2Vec2-Deepfake-Sentinel** project. It is specifically designed to train and evaluate robust audio deepfake detection models. By consolidating diverse spoofing techniques and high-quality authentic speech, this collection aims to provide a comprehensive defense mechanism against modern AI-generated voice fraud.

## 📊 Dataset Structure
The dataset is provided in four major compressed volumes to ensure integrity and ease of transfer:

| Filename | Description | Source/Notes |
| :--- | :--- | :--- |
| `LA.zip` | Logical Access (LA) subset | From ASVspoof 2019, covering various TTS and VC attacks. |
| `Fake_or_Real.zip` | Cross-domain Deepfake data | Mix of authentic human speech and high-fidelity deepfakes. |
| `In_The_Wild.zip` | Real-world spoofing scenarios | Audio collected from diverse, noisy, and unconstrained environments. |
| `WaveFake.zip` | Multi-generator spoofed speech | **Enhanced Version** (See Data Quality section). |

## 🛠️ Data Quality & Refinement (Crucial Updates)
To ensure the highest acoustic fidelity and avoid artifacts present in secondary processed versions, we have performed **Source-Level Data Replacement** within the `WaveFake` directory:

1. **JLSpeech Integration**: The original JLSpeech components in WaveFake have been replaced with the **official, high-quality JLSpeech source data** to ensure consistent sample rates and bit depths.
2. **JSUT (Japanese Speech) Refinement**: Replaced the Japanese subsets with the official **JSUT (Japanese Speech Corpus of Saruwatari Lab)** data, providing a more reliable baseline for multilingual deepfake detection.
3. **Integrity Check**: All audio files have been normalized to a consistent format suitable for Wav2Vec2 fine-tuning (e.g., 16kHz, Mono).

## 🚀 Usage for FinQuest 2026
This dataset is a private asset for the **2026 FinQuest Competition**.

### How to load
```python
from datasets import load_dataset

# Note: Access is restricted to authorized collaborators.
dataset = load_dataset("JesseHuang922/Wav2Vec2-Deepfake-Sentinel-Base-Dataset")