Voiceguard

🚀 Voice Detection AI - Real vs AI Audio Classifier

This model is an audio classifier based on fine - tuned Wav2Vec2. It can distinguish between real human voices and AI - generated voices. It has been trained on a dataset with samples from various TTS models and real human audio recordings.

🚀 Quick Start

To start using this model, you first need to install the necessary dependencies. Then you can load the model and processor, and perform audio classification.

✨ Features

• Accurate Classification: Capable of distinguishing between real human voices and AI - generated voices.
• Multiple Audio Formats Supported: Supports common audio formats such as .wav and .mp3.
• Robust Performance: Can classify across multiple AI - generation models.

📦 Installation

Make sure you have transformers and torch installed:

pip install transformers torch torchaudio

💻 Usage Examples

Basic Usage

import torch
from transformers import Wav2Vec2ForSequenceClassification, Wav2Vec2Processor
import torchaudio

# Load model and processor
model_name = "Mrkomiljon/voiceGUARD"
model = Wav2Vec2ForSequenceClassification.from_pretrained(model_name)
processor = Wav2Vec2Processor.from_pretrained(model_name)

# Load audio
waveform, sample_rate = torchaudio.load("path_to_audio_file.wav")

# Resample if necessary
if sample_rate != 16000:
    resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
    waveform = resampler(waveform)

# Preprocess
inputs = processor(waveform.squeeze().numpy(), sampling_rate=16000, return_tensors="pt", padding=True)

# Inference
with torch.no_grad():
    logits = model(**inputs).logits
    predicted_ids = torch.argmax(logits, dim=-1)

# Map to label
labels = ["Real Human Voice", "AI-generated"]
prediction = labels[predicted_ids.item()]
print(f"Prediction: {prediction}")

📚 Documentation

Model Overview

This model is a fine - tuned Wav2Vec2 - based audio classifier capable of distinguishing between real human voices and AI - generated voices. It has been trained on a dataset containing samples from various TTS models and real human audio recordings.

Model Details

Performance

• Robustness: Successfully classifies across multiple AI - generation models.
• Limitations: Struggles with certain unseen AI - generation models (e.g., ElevenLabs).

Training Procedure

• Data Collection: Compiled a balanced dataset of real human voices and AI - generated samples from various TTS models.
• Preprocessing: Standardized audio formats, resampled to 16 kHz, and adjusted durations to 10 seconds.
• Fine - Tuning: Utilized the Wav2Vec2 architecture for sequence classification, training for 3 epochs with a learning rate of 1e - 5.

Evaluation

• Metrics: Accuracy, Precision, Recall
• Results: Achieved 99.8% validation accuracy on the test set.

Limitations and Future Work

• While VoiceGUARD performs robustly across known AI - generation models, it may encounter challenges with novel or unseen models.
• Future work includes expanding the training dataset with samples from emerging TTS technologies to enhance generalization.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgements

• Special thanks to the developers of the Wav2Vec2 model and the contributors to the datasets used in this project.
• View the complete project on GitHub