AASIST3

🛡️ Speech Anti-Spoofing Arena

Independently re-scored on the reproducible Speech Anti-Spoofing Arena (EER %, lower is better; the model returns a score where higher = more bona fide):

EER% 9.44 on ASVspoof2019_LA EER% 28.73 on ASVspoof2021_DF EER% 29.72 on InTheWild EER% 30.73 on CD-ADD EER% 32.06 on ASVspoof2021_LA EER% 23.24 on SONAR EER% 31.82 on LibriSeVoc EER% 33.13 on CFAD EER% 44.23 on CVoiceFake_small EER% 34.59 on ASVspoof5 EER% 35.91 on DeepVoice EER% 36.91 on ArAD EER% 23.58 on DECRO EER% 16.16 on J-SPAW_LA EER% 39.04 on ODSS EER% 27.17 on HABLA EER% 1.4 on DFADD EER% 26.56 on PyAra EER% 28.84 on XMAD 1-SRR% 2.96 on LRLspoof EER% 24.69 on ADD22_eval_31 EER% 44.12 on ADD2023_track12_test_r1 EER% 7.31 on EmoFake_test 1-SRR% 18.61 on EmoSpoofTTS arena tier arena rank

Scores produced with the speech-spoof-bench wrapper: preemphasis (0.97) + a deterministic first-64,600-sample window; score = output logit for class 1 (bona fide). Pinned score files live under .eval_results/.

This repository contains the original implementation of AASIST3: KAN-Enhanced AASIST Speech Deepfake Detection using SSL Features and Additional Regularization for the ASVspoof 2024 Challenge.

Paper

AASIST3: KAN-Enhanced AASIST Speech Deepfake Detection using SSL Features and Additional Regularization for the ASVspoof 2024 Challenge

This is the original implementation of the paper. The model weights provided here are NOT the same weights used in the paper results.

Overview

AASIST3 is an enhanced version of the AASIST (Anti-spoofing with Adaptive Softmax and Instance-wise Temperature) architecture that incorporates Kolmogorov-Arnold Networks (KAN) for improved speech deepfake detection. The model leverages:

• Self-Supervised Learning (SSL) Features: Uses Wav2Vec2 encoder for robust audio representation
• KAN Linear Layers: Kolmogorov-Arnold Networks for enhanced feature transformation
• Graph Attention Networks (GAT): For spatial and temporal feature modeling
• Multi-branch Inference: Multiple inference branches for robust decision making

Architecture

The AASIST3 model consists of several key components:

1. Wav2Vec2 Encoder: Extracts SSL features from raw audio
2. KAN Bridge: Transforms SSL features using Kolmogorov-Arnold Networks
3. Residual Encoder: Processes features through multiple residual blocks
4. Graph Attention Networks:
   • GAT-S: Spatial attention mechanism
   • GAT-T: Temporal attention mechanism
5. Multi-branch Inference: Four parallel inference branches with master tokens
6. KAN Output Layer: Final classification using KAN linear layers

Key Innovations

• KAN Integration: Replaces traditional linear layers with KAN linear layers for better feature approximation
• Enhanced Regularization: Additional dropout and regularization techniques
• Multi-dataset Training: Trained on multiple ASVspoof datasets for robustness

🚀 Quick Start

Installation

git clone https://github.com/mtuciru/AASIST3.git
cd AASIST3
pip install -r requirements.txt

Loading the Model

from model import aasist3

# Load the model from Hugging Face Hub
model = aasist3.from_pretrained("MTUCI/AASIST3")
model.eval()

Basic Usage

import torch
import torchaudio

# Load and preprocess audio
audio, sr = torchaudio.load("audio_file.wav")
# Ensure audio is 16kHz and mono
if sr != 16000:
    audio = torchaudio.transforms.Resample(sr, 16000)(audio)
if audio.shape[0] > 1:
    audio = torch.mean(audio, dim=0, keepdim=True)

# Prepare input (model expects ~4 seconds of audio at 16kHz)
# Pad or truncate to 64600 samples
if audio.shape[1] < 64600:
    audio = torch.nn.functional.pad(audio, (0, 64600 - audio.shape[1]))
else:
    audio = audio[:, :64600]

# Run inference
with torch.no_grad():
    output = model(audio)
    probabilities = torch.softmax(output, dim=1)
    prediction = torch.argmax(probabilities, dim=1)
    
    # prediction: 0 = bonafide, 1 = spoof
    print(f"Prediction: {'Bonafide' if prediction.item() == 0 else 'Spoof'}")
    print(f"Confidence: {probabilities.max().item():.3f}")

Training Details

Datasets Used

The model was trained on a combination of multiple datasets:

• ASVspoof 2019 LA (Logical Access)
• ASVspoof 2024 (ASVspoof5)
• MLAAD (Multi-Language Audio Anti-Spoofing Dataset)
• M-AILABS (Multi-Language Audio Dataset)

Training Configuration

• Epochs: 20
• Batch Size: 12 (training), 24 (validation)
• Learning Rate: 1e-4
• Optimizer: AdamW
• Loss Function: CrossEntropyLoss
• Gradient Accumulation Steps: 2

Hardware

• GPUs: 2xA100 40GB
• Framework: PyTorch with Accelerate for distributed training

Advanced Usage

Custom Training

# Train the model
bash train.sh

Validation

# Run validation on test sets
bash validate.sh

Model Configuration

The model can be configured through the configs/train.yaml file:

# Key parameters
num_epochs: 20
train_batch_size: 12
val_batch_size: 24
learning_rate: 1e-4
gradient_accumulation_steps: 2

🤝 Citation

If you use this implementation in your research, please cite the original paper:

@inproceedings{borodin24_asvspoof,
  title     = {AASIST3: KAN-enhanced AASIST speech deepfake detection using SSL features and additional regularization for the ASVspoof 2024 Challenge},
  author    = {Kirill Borodin and Vasiliy Kudryavtsev and Dmitrii Korzh and Alexey Efimenko and Grach Mkrtchian and Mikhail Gorodnichev and Oleg Y. Rogov},
  year      = {2024},
  booktitle = {The Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024)},
  pages     = {48--55},
  doi       = {10.21437/ASVspoof.2024-8},
}

License

This project is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0) - see the LICENSE file for details.

This license allows you to:

• Share: Copy and redistribute the material in any medium or format
• Attribution: You must give appropriate...