voiceshield-aasist

VoiceShield — AASIST Universal Audio Deepfake Detector

Fine-tuned AASIST (Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks) for detecting deepfake/synthetic speech across multiple languages.

Model Details

Training Data

Fine-tuned on a merged dataset of:

• IndicTTS Deepfakes — Synthetic speech in Hindi, Tamil, Telugu, Bengali, Marathi, and other Indian languages
• ElevenLabs English Deepfakes — High-quality English AI-generated speech (garystafford dataset)

Original AASIST was pre-trained on ASVspoof2019 LA (Logical Access) dataset.

Usage

import torch
from huggingface_hub import hf_hub_download

# Download weights
weights_path = hf_hub_download(
    repo_id="Shauriya24/voiceshield-aasist",
    filename="AASIST_universal_best.pth"
)

# Load model (requires AASIST architecture code)
from models.AASIST import Model

model_config = {
    "architecture": "AASIST",
    "nb_samp": 64600,
    "first_conv": 128,
    "filts": [70, [1, 32], [32, 32], [32, 64], [64, 64]],
    "gat_dims": [64, 32],
    "pool_ratios": [0.5, 0.7, 0.5, 0.5],
    "temperatures": [2.0, 2.0, 100.0, 100.0]
}

model = Model(model_config)
model.load_state_dict(torch.load(weights_path, map_location="cpu"))
model.eval()

Citation

@inproceedings{Jung2022AASIST,
  title={AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks},
  author={Jee-weon Jung and Hee-Soo Heo and Hemlata Tak and others},
  booktitle={ICASSP},
  year={2022}
}

Acknowledgements

• Base architecture by NAVER Corp (MIT License)
• Fine-tuned as part of the VoiceShield project for Smart India Hackathon (SIH)