bass

Code for Bayesian auditory scene synthesis (project page).

The corresponding publication is:

Maddie Cusimano, Luke B. Hewitt, Josh H. McDermott. (2024). Listening with generative models. Cognition. 253, 105874. doi: https://doi.org/10.1016/j.cognition.2024.105874.

The READMEs in this repository, as well as the code, contain references to sections, equations and figures, which correspond to the numbering in this publication.

Directory overview

• model/: The generative model. See sections 2.2.1 and 2.2.2, and Appendix A.
• renderer/: The renderer/synthesizer which transforms a scene description into a sound. See section 2.2.3 and section A.5.
• inference/: enumerative and sequential inference algorithms (with a lot of slurm-specific code). See section 2.3 and Appendix B.
• psychophysics/: Code for experiments with classic illusions, see Appendix C.
  • generation/: Generation of stimuli
  • hypotheses/: Creation of initial hypotheses for enumerative inference
  • analysis/: Analysis of model inferences to get experiment results, and plot creation
  • comparisons/: Code to compute human-model dissimilarity, see section 2.6 and Appendix D.
• everyday_sound_expts/: Code to create and analyze everyday sound experiments, see section 2.8 and Appendix F.

Notes

This code is provided as a reference for the paper to help demonstrate the logic behind the model and experiments, rather than a fully executable package.

Citation

@article{cusimano2024listening,
	title = {Listening with generative models},
	journal = {Cognition},
	volume = {253},
	pages = {105874},
	year = {2024},
	issn = {0010-0277},
	doi = {https://doi.org/10.1016/j.cognition.2024.105874},
	url = {https://www.sciencedirect.com/science/article/pii/S0010027724001604},
	author = {Maddie Cusimano and Luke B. Hewitt and Josh H. McDermott},
}