Data from : Evaluating and Predicting the Audibility of Acoustic Alarms in the Workplace Using Experimental Methods and Deep Learning

Description

This repository serves as a complementary resource accompanying the academic paper titled "Evaluating and Predicting the Audibility of Acoustic Alarms in the Workplace Using Experimental Methods and Deep Learning" published in Applied Acoustics (available at: https://doi.org/10.1016/j.apacoust.2024.109955). It comprises a dataset containing the acoustic data, metadata, and perceptual annotations.

In addition, we provide the .h5 datasets and PyTorch model weights to run the code corresponding to the neural network section discussed in the paper (publicly accessible at: https://github.com/effajr/predicting_alarm_audibility).

The repository is composed of five .zip files:

source_audio : contains the source audio files used to generate the alarms and backgrounds present in the data file.

data : contains the audio files corresponding to the alarms and backgrounds, along with the perceptual annotations.

metadata : contains the metadata related to the alarms and backgrounds present in the data file, and to the source audio files contained in source_audio.

features : contains the .h5 files representing the development and evaluation subsets (mel-spectrograms and perceptual labels) used in the deep learning approach presented in the paper.

trained_models : contains the PyTorch model weights for the 10 runs of model training mentionned in the paper.

How to use the data

To run the code present in the GitHub repository, we recommend extracting the files data.zip, features.zip and trained_models.zip in their corresponding folders in the "application" folder (see https://github.com/effajr/predicting_alarm_audibility).

Content

Content of data.zip 

annotations/
├─ dev/
│ ├─ annotation_compilation_dev.csv : Compilation of all the listening conditions and │  │ individual annotator responses for the development data.│ ├─ dev_conditions.csv : Unique listening conditions (extracted from annotation_compilation_dev.csv).
│ ├─ dev_labels.csv : All individual annotator responses for each │ │ unique listening condition (extracted from annotation_compilation_dev.csv).
│ ├─ dev_train_valid_split.csv : Random 80%/20% training/validation split used for development │ │ in the experiments reported in the paper.
├─ eval/│ ├─ annotation_compilation_eval.csv : Compilation of all the listening conditions and individual annotator │  │                                     responses for the evaluation data.│  │                                     The column 'clearly_audible_mean' represents individual annotator │ │                                    binary responses evaluated for each listening condition.│  │                                     The column 'clearly_audible_pf' represents individual annotator │  │                                     psychometric functions evaluated for each listening condition.│  ││ ├─ eval_conditions.csv : Unique listening conditions (extracted from annotation_compilation_eval.csv).
│ ├─ eval_labels_apf.csv : All individual annotator psychometric function values for each │  │ unique listening condition (extracted from annotation_compilation_eval.csv).
│ ├─ eval_labels_mv.csv : All individual annotator binary responses for each │  │ unique listening condition (extracted from annotation_compilation_eval.csv)│audio/ : .wav files corresponding to the alarms and backgrounds for Development and Evaluation subsets of the dataset.
├─ dev/
│ ├─ alarms/
│ ├─ backgrounds/
├─ eval/
│ ├─ alarms/
│ ├─ backgrounds/Content of metadata.zip ├─ audio_metadata.xlsx : Table of the alarms and background files with short descriptions, │ source file names, and temporal information (in seconds).
├─ source_file_metadata.xlsx : Metadata table of the original files used to generate alarms and backgrounds.