Description

This repository provides the dataset and demo code for AuralNet.

Dataset

• Download the dataset here. Place the downloaded files in the ./dataset directory.
• This dataset contains extracted features and ground truth labels of AuralNet, including both MCT and clean training and testing sets.
• The data is in MATLAB v7.3 format.e
  • The feature dimension is sample size * 3, with each column representing the log-Gammatone power spectrogram of the left ear (64 * 39), the log-Gammatone power spectrogram of the right ear (64 * 39), and cross-correlation values (33 * 1).
  • The label dimension is sample size * 8 * 3, where 8 represents 8 sectors, and 3 represents labels of sound source presence (1 for present, 0 for absence), normalized azimuth (1 for no source), and normalized elevation (-1 for no source).

Files

• model.py: Demo code for AuralNet, implemented with TensorFlow 2.5.0.
  • Requires mat73 for loading MATLAB v7.3 files.
• ./model_save/MCT_weights.h5: Pretrained model weights.
• ./dataset: Directory for storing the dataset.

Testing

To test the pretrained MCT model, run the following command:

python test.py

Training a New Model

To train a new MCT model from scratch, use the following command:

python train.py

Additional Notes

• Ensure that you have installed all the required dependencies before running the test or training scripts.
• If you encounter any issues, please contact linya.fu@outlook.com