README.md

---
license: apache-2.0
datasets:
- 012shin/fake-audio-detection-augmented
language:
- en
metrics:
- accuracy
- f1
- recall
- precision
base_model:
- MIT/ast-finetuned-audioset-10-10-0.4593
pipeline_tag: audio-classification
library_name: transformers
tags:
- audio
- audio-classification
- fake-audio-detection
- ast
widget:
  - text: "Upload an audio file to check if it's real or synthetic"
inference:
  parameters:
    sampling_rate: 16000
    audio_channel: "mono"
model-index:
- name: ast-fakeaudio-detector
  results:
  - task:
      type: audio-classification
      name: Audio Classification
    dataset:
      name: fake-audio-detection-augmented
      type: 012shin/fake-audio-detection-augmented
    metrics:
      - type: accuracy
        value: 0.9662
      - type: f1
        value: 0.9710
      - type: precision
        value: 0.9692
      - type: recall
        value: 0.9728
---

# AST Fine-tuned for Fake Audio Detection

This model is a binary classification head fine-tuned version of [MIT/ast-finetuned-audioset-10-10-0.4593](https://huggingface.co/MIT/ast-finetuned-audioset-10-10-0.4593) for detecting fake/synthetic audio. The original AST (Audio Spectrogram Transformer) classification head was replaced with a binary classification layer optimized for fake audio detection.

## Model Description

- **Base Model**: MIT/ast-finetuned-audioset-10-10-0.4593 (AST pretrained on AudioSet)
- **Task**: Binary classification (fake/real audio detection)
- **Input**: Audio converted to Mel spectrogram (128 mel bins, 1024 time frames)
- **Output**: Binary prediction (0: real audio, 1: fake audio)
- **Training Hardware**: 2x NVIDIA T4 GPUs

## Training Configuration

```python
{
    'learning_rate': 1e-5,
    'weight_decay': 0.01,
    'n_iterations': 1500,
    'batch_size': 16,
    'gradient_accumulation_steps': 8,
    'validate_every': 500,
    'val_samples': 5000
}
```

## Dataset Distribution

The model was trained on a filtered dataset with the following class distribution:

```
Training Set:
- Fake Audio (0): 29,089 samples (53.97%)
- Real Audio (1): 24,813 samples (46.03%)

Test Set:
- Fake Audio (0): 7,229 samples (53.64%)
- Real Audio (1): 6,247 samples (46.36%)
```

## Model Performance

Final metrics on validation set:
- Accuracy: 0.9662 (96.62%)
- F1 Score: 0.9710 (97.10%)
- Precision: 0.9692 (96.92%)
- Recall: 0.9728 (97.28%)

# Usage Guide

## Model Usage
```python
import torch
import torchaudio
import soundfile as sf
import numpy as np
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification

# Load model and move to available device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_name = "WpythonW/ast-fakeaudio-detector"

extractor = AutoFeatureExtractor.from_pretrained(model_name)
model = AutoModelForAudioClassification.from_pretrained(model_name).to(device)
model.eval()

# Process multiple audio files
audio_files = ["audio1.wav", "audio2.mp3", "audio3.ogg"]
processed_batch = []

for audio_path in audio_files:
    # Load audio file
    audio_data, sr = sf.read(audio_path)
    
    # Convert stereo to mono if needed
    if len(audio_data.shape) > 1 and audio_data.shape[1] > 1:
        audio_data = np.mean(audio_data, axis=1)
    
    # Resample to 16kHz if needed
    if sr != 16000:
        waveform = torch.from_numpy(audio_data).float()
        if len(waveform.shape) == 1:
            waveform = waveform.unsqueeze(0)
        
        resample = torchaudio.transforms.Resample(
            orig_freq=sr, 
            new_freq=16000
        )
        waveform = resample(waveform)
        audio_data = waveform.squeeze().numpy()
    
    processed_batch.append(audio_data)

# Prepare batch input
inputs = extractor(
    processed_batch,
    sampling_rate=16000,
    padding=True,
    return_tensors="pt"
)
inputs = {k: v.to(device) for k, v in inputs.items()}

# Get predictions
with torch.no_grad():
    logits = model(**inputs).logits
    probabilities = torch.nn.functional.softmax(logits, dim=-1)

# Process results
for filename, probs in zip(audio_files, probabilities):
    fake_prob = float(probs[0].cpu())
    real_prob = float(probs[1].cpu())
    prediction = "FAKE" if fake_prob > real_prob else "REAL"
    
    print(f"\nFile: {filename}")
    print(f"Fake probability: {fake_prob:.2%}")
    print(f"Real probability: {real_prob:.2%}")
    print(f"Verdict: {prediction}")
```

## Limitations

Important considerations when using this model:
1. The model works best with 16kHz audio input
2. Performance may vary with different types of audio manipulation not present in training data
3. Very short audio clips (<1 second) might not provide reliable results
4. The model should not be used as the sole determiner for real/fake audio detection

## Training Details

The training process involved:
1. Loading the base AST model pretrained on AudioSet
2. Replacing the classification head with a binary classifier
3. Fine-tuning on the fake audio detection dataset for 1500 iterations
4. Using gradient accumulation (8 steps) with batch size 16
5. Implementing validation checks every 500 steps