DNN-binaural-localization

1. Simulate directional sound with head related impulse responses (HRIRs)
2. Train a deep neural network (DNN) to locate sound sources
3. Evaluate the DNN and analyse it with layer-wise relevance propagation (LRP)

This repository contains the source code of my Bachelor thesis with the topic "An Analysis of Binaural Localization Models with Explainable Machine Learning Methods". A PDF version is available on my website: http://pretzelsolution.com/binaural-localization

Code structure

• 1_data_presentation.ipynb: Visually investigation of the HRIR database, directional sound simulation and decision of which preprocessing is applied for the input feature vector. Interaural level difference (ILD) has been chosen.
• 2_training.ipynb: DNN training with grid search hyperparameter optimization
• 3_evaluation.ipynb: Evaluate DNNs with different error metrics and explain its predictions with LRP
• custom_dataset_class.ipynb: Additional dataset class which can be used to conduct memory efficiant preprocessing in batches

Quickstart

1. Create a folder named with "binaural_localization" in your Google Drive
2. Copy content of this repository into the folder
3. Doube-click on a Jupiter notebook, e.g. 1_data_presentation.ipynb, and choose "open with" "Google colab"
4. Mount Google Drive
5. Run the code

License

Content in the folder LRP is licensed under the BSD License 2.0 by Moritz Böhle and available in this repository.

The HRIR database was created by J. Arend et. al. under Attribution-ShareAlike 3.0 Unported and can be downloaded here