Voice Fake Detector v1
This model detects whether an audio clip is real human speech or AI-generated (deepfake). It is based on a fine-tuned Wav2Vec2 architecture and designed for robust audio deepfake detection.
Overview
The model performs binary classification:
- Real (human voice)
- Fake (AI-generated voice)
It is suitable for deepfake detection, voice verification, and media authenticity applications.
Model Details
- Task: Audio Classification
- Architecture: Wav2Vec2 (Transformer-based)
- Base Model: facebook/wav2vec2-xls-r-300m
- Framework: PyTorch (Transformers)
- Input:
.wavaudio - Output: Label with confidence score
Usage
from transformers import pipeline
classifier = pipeline("audio-classification", model="Shanmugapriya6/voice-fake-detector-v1")
result = classifier("audio.wav")
print(result)
Example
Input: audio.wav Output: Fake (confidence: 0.91)
Training Data
The model was trained on a combination of real and synthetic speech data.
Real Speech Datasets
- Svarah dataset (AI4Bharat)
- Kathbath dataset (AI4Bharat) – Tamil subset accessed via AIKosh platform
- Indian Languages Audio Dataset (Kaggle): https://www.kaggle.com/datasets/hmsolanki/indian-languages-audio-dataset
Synthetic (Fake) Audio
- AI-generated speech samples created using text-to-speech (TTS) systems
Preprocessing
- Audio resampling
- Silence trimming
- Normalization
- Temporal chunking
Performance
Evaluation results on ASVspoof2019 subset:
- Accuracy: 0.9286
- Precision: 0.9999
- Recall: 0.9205
- F1 Score: 0.9363
- Equal Error Rate (EER): 0.0401
Note: Performance may vary depending on dataset and audio conditions.
Limitations
- Performance may degrade on noisy or low-quality audio
- Not evaluated across all languages and accents
- May not detect highly advanced deepfake techniques
Out-of-Scope Use
- Not intended for legal or forensic decisions
- Not recommended for high-risk authentication systems
License
MIT