A benchmark comparison of perceptual models for soundscapes on a large-scale augmented soundscape dataset
Psychoacoustic indicators and spectrogram-based features are standard inputs to perceptual models for soundscape analysis. However, existing models in the soundscape literature are trained on different collections of input parameters and frequently mutually exclusive datasets, which complicates fair...
Saved in:
Main Authors: | , , , , |
---|---|
Other Authors: | |
Format: | Conference or Workshop Item |
Language: | English |
Published: |
2023
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/164983 https://www.ica2022korea.org/data/Proceedings_A14.pdf https://ica2022korea.org/ |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Psychoacoustic indicators and spectrogram-based features are standard inputs to perceptual models for soundscape analysis. However, existing models in the soundscape literature are trained on different collections of input parameters and frequently mutually exclusive datasets, which complicates fair comparisons of model performance and generalizability to unseen data and soundscapes. Hence, we use the ARAUS dataset, a large-scale, publicly-available dataset of 25,440 responses to unique augmented soundscapes, as a common benchmark for comparison of the relative performance of a selection of input and model types used in previous soundscape studies, as well as deep neural network architectures commonly used for other acoustic tasks. The different model types were used in a regression task to predict “Eventfulness” ratings (as defined in ISO/TS 12913-3) given by participants as responses to the augmented soundscapes. Subsequently, we compared their performance as classification models for the classification of soundscapes into the quadrants generated by the Pleasantness-Eventfulness axes of the ISO/TS 12913-3 circumplex model. The five-fold cross-validation set of 25,200 responses and an independent test set of 240 responses making up the ARAUS dataset was used to facilitate unbiased comparisons. |
---|