LibriVAD Dataset Generation

This repository contains the scripts necessary to generate the LibriVAD dataset, a large-scale, noise-augmented dataset for Voice Activity Detection (VAD) based on the LibriSpeech corpus.

Setup LibriVAD Requirements

1. Install Python Libraries: Navigate to the LibriVAD directory and install the required dependencies.

   pip install -r requirements.txt

2. Download Source Data: Run the setup.py script and choose the LibriSpeech mirror closest to your location from [EU, USA, CN]. This will download the necessary splits of the LibriSpeech dataset, as well as the Forced_alignments and Noises data from Hugging Face.

   EU Example:

   python setup.py EU

   After downloading train-clean-100, dev-clean, and test-clean, the local setup should take a few minutes.

Creating the LibriVAD Dataset

3. Generate the Dataset: After the setup is complete, run the create_LibriVAD.py script. Choose the size of the generated dataset: "small", "medium", or "large".

   The approximate sizes of the final dataset are 15GB, 150GB, and 1.5TB respectively.

   Small Dataset Example:

   python create_LibriVAD.py small

   The generation process can take several hours, depending on the chosen size and your machine's performance.

Warning: The initial space requirement for the setup process (before generating the final dataset) is approximately 61GB. Please ensure you have sufficient disk space.

Noisy Signals

The noise used for the generation of LibriVAD can be downloaded from https://huggingface.co/datasets/LibriVAD/LibriVAD/resolve/main/Files/Noises.zip

The trained models can also be found in the Hugging Face page: https://huggingface.co/datasets/LibriVAD/LibriVAD/tree/main/Eval-ViT-MFCC

License

This project is licensed under the MIT License.

Citation

I. Stylianou, A.K. Sarkar, N. Dawalatabad, J. Glass, and Z.-H. Tan, "LibriVAD: A Scalable Open Dataset with Deep Learning Benchmarks for Voice Activity Detection," arXiv preprint arXiv:2512.17281 (2025). url