Synthetic Speech Detection using the Fake-or-Real (FoR) Dataset

Project Overview

This project aims to develop a deep learning model to detect synthetic speech using the Fake-or-Real (FoR) dataset. The model distinguishes between real human speech and computer-generated speech, addressing the growing concern of deepfake audio. We focus on the for-norm version of the dataset.

Dataset

The Fake-or-Real (FoR) dataset is a comprehensive collection of over 195,000 utterances, including both real human speech and computer-generated speech. It incorporates data from:

• Latest TTS solutions (e.g., Deep Voice 3, Google Wavenet TTS)
• Arctic Dataset
• LJSpeech Dataset
• VoxForge Dataset
• Custom speech recordings

Dataset Version

This project uses the for-norm version of the dataset, which is balanced for gender and class, and normalized for sample rate, volume, and number of channels.

Dataset Source

The dataset is downloaded from Kaggle: The Fake-or-Real Dataset (for-norm version)

Project Structure

audio_classification_project/
│
├── data/
│   ├── train/
│   │   ├── real/
│   │   └── fake/
│   └── test/
│       ├── real/
│       └── fake/
│
├── models.py
├── dataloader.py
├── train.py
├── eval.py
├── utils.py
├── metrics.py
├── main.py
├── run.sh
├── requirements.txt
└── README.md

Setup and Installation

1. Clone the repository:

   git clone https://github.com/himudigonda/FoR_Audio
   cd FoR_Audio
   

2. Create a virtual environment:

   mamba create -n FoR_Audio python==3.10
   mamba activate FoR_Audio
   

3. Install the required packages:

   pip install -r requirements.txt
   

4. Download the for-norm dataset from the Kaggle link provided above and place the audio files in the appropriate directories under data/.

Usage

1. To train the model:

   python main.py
   

2. To evaluate the model:

   python eval.py
   

Alternatively, you can use the provided shell script to run both training and evaluation:

bash run.sh

Model Architecture

The project uses a convolutional neural network (CNN) to classify mel spectrograms of audio samples. The architecture is defined in models.py.

Results

(Add information about model performance, accuracy, and any notable findings once you have results)

Contributing

Contributions to improve the model or extend the project are welcome. Please feel free to submit a Pull Request.

License

(Add appropriate license information)

Acknowledgements

• The Fake-or-Real (FoR) Dataset creators
• APTLY and LaSSoftE, Lassonde School of Engineering
• Kaggle for hosting the dataset

Citations:

[1] https://www.kaggle.com/datasets/mohammedabdeldayem/the-fake-or-real-dataset?select=for-norm

[2] https://www.kaggle.com/docs/datasets

[3] https://www.kaggle.com/discussions/general/156610

[4] https://www.youtube.com/watch?v=W86uvkzaqLg

[5] https://setronica.com/how-to-use-kaggle-datasets-for-research-a-step-by-step-guide/

[6] https://www.datacamp.com/tutorial/tutorial-kaggle-datasets-tutorials-kaggle-notebooks

[8] https://mostly.ai/blog/ml-classifier-for-fake-vs-real-data

[9] https://mostly.ai/video/fake-or-real-how-to-use-a-data-discriminator-for-evaluating-synthetic-data-quality