---
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%]