Estimation of Direction of Arrival (DOA) for First Order Ambisonic Audio Files using Artificial Neural Networks

Pedro Pablo Lucas Bravo

University of Oslo (UiO), Norwegian University of Science and Technology (NTNU)

Music, Communication and Technology Master Programme

[email protected]

I am registered on MCT4047 UiO's Studentweb

In the notebook file.ipybn you will find this content. You can navigate to other notebooks from there.

The aim of this project is to develop a solution to estimate the Direction of Arrival(DOA) from a sound event encoded as a First Order Ambisonic audio file (4-channel). The source code is organized in several Jupyter notebooks that are linked to the instructions given below.

For executing the project follow the next steps (After downloading the data, you can executed any of them if the focus is in one of these stages). Every notebook has an approximate execution time provided below:

1. Download the data from here, uncompress the folder at the level of this file.ipynb. You should be able to have a new folder called data that contains another folder called dcase_data.

2. Download the extracted features and model from here, uncompress at the level of this file.ipynb. You should be able to have 3 files: features_train.csv, features_test.csv, doa_model.pkl. (Note that this step is not necessary if the intention is to extract the features and save as the next step indicates).

3. Execute the notebook feature_extraction.ipybn. Here you can generate the features for training and testing and save them to files if needed. (Exec. time 30 min)

4. Execute the notebook training.ipybn. In this notebook the features for training will be loaded from one of the files generated in the previous step. Here the models will be saved if needed. (Exec. time 7 min)

5. Execute the notebook inference.ipybn. The data from the features for testing will be loaded to show performance metrics related with instances that the ML algorithm do not consider in the training process. (Exec. time 15 sec). This notebook uses the library SoundFile to save part of the samples in FOA(4-channel) .wav files. In case you don't have it, include it in your environment by running the next cell for installation (This is found in the file.ipybn notebook).

Experiments regarding feature extraction and training can be found in the folder experiments. They will NOT run since they are not in the root folder, but they show part of the effort to find the right features and network architecture.