AnodHuang/AST-AMVD-SAD-v2

AST-AMVD-SAD-v1

Description

A fine-tuned audio classification model for detecting AI-generated audio content.

Author

• Kunyang Huang ([email protected])
• Bin Hu ([email protected])

Model Details

Model Description

• Architecture: Based on the Audio Spectrogram Transformer (AST) architecture from MattyB95/AST-ASVspoof2019-Synthetic-Voice-Detection
• Input: Audio waveforms converted to mel-spectrogram representations
• Output: Four-class classification for audio authenticity detection

Intended Use

This model is designed to:

• Detect AI-generated audio content
• Identify different types of synthetic audio:
  • Class 0 (H): Real Human Audio
  • Class 1 (C): AI Cloned Audio
  • Class 2 (A): AI Generated Audio
  • Class 3 (Combined): Mixed Human/AI Audio
• Primary use cases include:
  • Content authenticity verification
  • AI-generated content detection systems
  • Audio forensics applications

Training Data

• Dataset: AMVD_AS Dataset
• Data Composition:
  • Balanced samples across four categories
  • Contains both synthetic and genuine human audio samples

Training Procedure

Fine-tuning Parameters

• Base Model: MattyB95/AST-ASVspoof2019-Synthetic-Voice-Detection
• Initial Learning Rate: 4e-5 → 1e-5 (linear decay)
• Total Training Steps: 25,000
• Batch Size: 32
• Warmup Steps: 5,000
• Weight Decay: 0.01
• Gradient Clip Norm: 1.0
• Training Duration: ~4.5 hours (A100 GPU)

Evaluation

Validation Performance

• Training Loss: 0.0874
• Eval Loss: 0.07367
• Eval Accuracy: 0.98109
• Final Steps per Second: 2.566
• Final Samples per Second: 10.264
• Runtime at 25k Steps: 824.1802
• Gradient Norm: 0.000075778
• LR Stability: 1e-5