Problems using deep generative models for probabilistic audio source separation

Maurice Frank, Maximilian Ilse
Proceedings on "I Can't Believe It's Not Better!" at NeurIPS Workshops, PMLR 137:53-59, 2020.

Abstract

Recent advancements in deep generative modeling make it possible to learn prior distributions from complex data that subsequently can be used for Bayesian inference. However, we find that distributions learned by deep generative models for audio signals do not exhibit the right properties that are necessary for tasks like audio source separation using a probabilistic approach. We observe that the learned prior distributions are either discriminative and extremely peaked or smooth and non-discriminative. We quantify this behavior for two types of deep generative models on two audio datasets.

Cite this Paper

BibTeX
@InProceedings{pmlr-v137-frank20a, title = {Problems using deep generative models for probabilistic audio source separation}, author = {Frank, Maurice and Ilse, Maximilian}, booktitle = {Proceedings on "I Can't Believe It's Not Better!" at NeurIPS Workshops}, pages = {53--59}, year = {2020}, editor = {Zosa Forde, Jessica and Ruiz, Francisco and Pradier, Melanie F. and Schein, Aaron}, volume = {137}, series = {Proceedings of Machine Learning Research}, month = {12 Dec}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v137/frank20a/frank20a.pdf}, url = {https://proceedings.mlr.press/v137/frank20a.html}, abstract = {Recent advancements in deep generative modeling make it possible to learn prior distributions from complex data that subsequently can be used for Bayesian inference. However, we find that distributions learned by deep generative models for audio signals do not exhibit the right properties that are necessary for tasks like audio source separation using a probabilistic approach. We observe that the learned prior distributions are either discriminative and extremely peaked or smooth and non-discriminative. We quantify this behavior for two types of deep generative models on two audio datasets.} }
Endnote
%0 Conference Paper %T Problems using deep generative models for probabilistic audio source separation %A Maurice Frank %A Maximilian Ilse %B Proceedings on "I Can't Believe It's Not Better!" at NeurIPS Workshops %C Proceedings of Machine Learning Research %D 2020 %E Jessica Zosa Forde %E Francisco Ruiz %E Melanie F. Pradier %E Aaron Schein %F pmlr-v137-frank20a %I PMLR %P 53--59 %U https://proceedings.mlr.press/v137/frank20a.html %V 137 %X Recent advancements in deep generative modeling make it possible to learn prior distributions from complex data that subsequently can be used for Bayesian inference. However, we find that distributions learned by deep generative models for audio signals do not exhibit the right properties that are necessary for tasks like audio source separation using a probabilistic approach. We observe that the learned prior distributions are either discriminative and extremely peaked or smooth and non-discriminative. We quantify this behavior for two types of deep generative models on two audio datasets.
APA
Frank, M. & Ilse, M.. (2020). Problems using deep generative models for probabilistic audio source separation. Proceedings on "I Can't Believe It's Not Better!" at NeurIPS Workshops, in Proceedings of Machine Learning Research 137:53-59 Available from https://proceedings.mlr.press/v137/frank20a.html.

Related Material