|
|
---
|
|
|
license: apache-2.0
|
|
|
tags:
|
|
|
- audio-classification
|
|
|
- deep-speech-detection
|
|
|
- tensorflow
|
|
|
- keras
|
|
|
---
|
|
|
# Model Card for Deep Speech Detection
|
|
|
|
|
|
## Model Description
|
|
|
This is a TensorFlow/Keras CNN model trained to detect deepfake or synthetic speech with >95% accuracy. It uses audio features (MFCCs, chroma, spectral centroid, etc.) extracted with `librosa`.
|
|
|
|
|
|
## Intended Use
|
|
|
- Deepfake speech detection
|
|
|
- Audio authenticity verification
|
|
|
|
|
|
## Dependencies
|
|
|
```bash
|
|
|
pip install tensorflow==2.10.0 librosa==0.10.1 joblib==1.3.2 numpy==1.22.4 pandas==1.5.3 scikit-learn==1.2.2
|
|
|
```
|
|
|
|
|
|
## Usage
|
|
|
```python
|
|
|
import tensorflow as tf
|
|
|
import librosa
|
|
|
import joblib
|
|
|
import numpy as np
|
|
|
import pandas as pd
|
|
|
from huggingface_hub import hf_hub_download, HfApi
|
|
|
import os
|
|
|
|
|
|
# Download model and files
|
|
|
repo_name = "Prince53/deep-speech-detection"
|
|
|
model_dir = "downloaded_model"
|
|
|
scaler_path = hf_hub_download(repo_name, "scaler.pkl", local_dir=model_dir)
|
|
|
label_encoder_path = hf_hub_download(repo_name, "label_encoder.pkl", local_dir=model_dir)
|
|
|
api = HfApi()
|
|
|
api.snapshot_download(repo_name, local_dir=model_dir, allow_patterns="saved_model/*")
|
|
|
|
|
|
# Load model and preprocessing objects
|
|
|
model = tf.keras.models.load_model(os.path.join(model_dir, "saved_model"))
|
|
|
scaler = joblib.load(scaler_path)
|
|
|
label_encoder = joblib.load(label_encoder_path)
|
|
|
|
|
|
# Feature extraction function
|
|
|
def segment_and_extract_features(audio, sr=16000):
|
|
|
segment_samples = int(2.0 * sr)
|
|
|
step_samples = int(0.25 * sr)
|
|
|
segments = [audio[i:i+segment_samples] for i in range(0, len(audio) - segment_samples + 1, step_samples)]
|
|
|
features = []
|
|
|
for segment in segments:
|
|
|
if len(segment) < segment_samples:
|
|
|
continue
|
|
|
mfccs = librosa.feature.mfcc(y=segment, sr=sr, n_mfcc=13)
|
|
|
chroma = librosa.feature.chroma_stft(y=segment, sr=sr)
|
|
|
spectral_centroid = librosa.feature.spectral_centroid(y=segment, sr=sr)
|
|
|
spectral_bandwidth = librosa.feature.spectral_bandwidth(y=segment, sr=sr)
|
|
|
rolloff = librosa.feature.spectral_rolloff(y=segment, sr=sr)
|
|
|
zero_crossing_rate = librosa.feature.zero_crossing_rate(y=segment)
|
|
|
feature_dict = {
|
|
|
'mfcc_mean': np.mean(mfccs, axis=1),
|
|
|
'mfcc_std': np.std(mfccs, axis=1),
|
|
|
'chroma': np.mean(chroma, axis=1),
|
|
|
'spectral_centroid': np.mean(spectral_centroid),
|
|
|
'spectral_bandwidth': np.mean(spectral_bandwidth),
|
|
|
'rolloff': np.mean(rolloff),
|
|
|
'zero_crossing_rate': np.mean(zero_crossing_rate)
|
|
|
}
|
|
|
features.append(feature_dict)
|
|
|
return features
|
|
|
|
|
|
# Classify audio
|
|
|
audio, sr = librosa.load("path/to/audio.wav", sr=16000)
|
|
|
segments = segment_and_extract_features(audio, sr)
|
|
|
segment_features = pd.concat([
|
|
|
pd.DataFrame([seg['mfcc_mean'] for seg in segments]),
|
|
|
pd.DataFrame([seg['mfcc_std'] for seg in segments]),
|
|
|
pd.DataFrame([seg['chroma'] for seg in segments]),
|
|
|
pd.DataFrame([[seg['spectral_centroid'], seg['spectral_bandwidth'], seg['rolloff'], seg['zero_crossing_rate']] for seg in segments])
|
|
|
], axis=1)
|
|
|
segment_features = scaler.transform(segment_features)
|
|
|
segment_features = segment_features.reshape(segment_features.shape[0], segment_features.shape[1], 1)
|
|
|
predictions = model.predict(segment_features)
|
|
|
segment_labels = np.argmax(predictions, axis=1)
|
|
|
confidence_scores = np.mean(predictions, axis=0)
|
|
|
final_label = label_encoder.inverse_transform([np.argmax(np.bincount(segment_labels))])[0]
|
|
|
print(f"Confidence Scores: Real={confidence_scores[0]:.4f}, Fake={confidence_scores[1]:.4f}")
|
|
|
print(f"Classification: {final_label} ({0 if final_label == 'Real' else 1})")
|
|
|
```
|
|
|
|
|
|
## Limitations
|
|
|
- Requires mono audio at 16kHz sampling rate.
|
|
|
- May struggle with low-quality audio or unseen domains.
|
|
|
- Trained on the Comb4 dataset.
|
|
|
|
|
|
## Training Data
|
|
|
- Dataset: Comb4 (custom dataset with real and fake audio)
|
|
|
- Size: [Update with number of samples]
|
|
|
|
|
|
## Evaluation
|
|
|
- Test Accuracy: [Update with >95%]
|
|
|
|
