article / 3 October 2022

Audio Across Domains Workshop 2022

A collaborative and cross-disciplinary meeting of audio data scientists spawns creative research collaborations 

In August, the Kitzes Lab at the University of Pittsburgh hosted the first AudioXD (“Audio Across Domains”) workshop. The Academic Data Science Alliance and the Gordon and Betty Moore Foundation provided support to host this three day cross-disciplinary meeting of researchers who work with audio data and audio data analysis. About 40 researchers attended, representing a variety of disciplines including ecology, linguistics, business, music, computer science, and others. 

Probably of most interest to the WILDLABS community, bioacoustics researchers were well-represented at the meeting, making up about two-thirds of attendees. These attendees mainly use field audio recording, mostly terrestrial, to answer questions in ecology and conservation related to species’ distributions, richness, and abundance. The group included specialists in many parts of the bioacoustics pipeline, including machine learning researchers, developers of infrastructure and platforms for serving audio data, and representatives from government agencies interested in leveraging bioacoustics data to understand and protect wildlife.

AudioXD was a participant-led “unconference,” meaning that attendees collaboratively created an agenda, drove discussion on topics of interest, and formed collaborations to continue projects going forward. We thought that there were four action items of particular interest that emerged from the workshop: creating interoperable systems, listing available data and models, continuing new frontiers in data analysis methods, and meeting again to continue to strengthen connection and collaboration within the discipline.

Creating interoperable systems

It became clear very quickly that there is unlikely to be a “one-size-fits-all” software platform that will meet the needs of every researcher, so attendees identified the importance of making platforms interoperable. Several attendees shared their experiences creating platforms for ingesting, managing, and displaying bioacoustic data. The features needed by attendees varied widely, and some had restrictions that made using existing platforms impossible. Participants also mentioned that platform diversity enables specialization and spurs experimentation. 

Although some diversity of audio software platforms may be inevitable or even desirable, these platforms do need to “speak the same language” so that users can port their data and analyses between different platforms. For example, if users can train or use machine learning classifiers in one platform, they should be able to export trained models or import and apply their own models. The group stressed the desirability of standardized, interoperable formats for data annotation and metadata as well.

Listing available data and models

Attendees suggested creating a central database of field data, annotated data, and models to facilitate reuse and collaboration. This database would not act as an archive for data or models: in many cases, researchers may not be able or ready to share all of their data freely. Instead, it could simply list previous or ongoing work and identify how researchers can request more information on a particular dataset or model. This would increase data discoverability and facilitate future partnerships.

Frontiers in data analysis methods

The group discussed that methods for species-level classification are becoming somewhat routine, and focused on discussions of two frontiers in data analysis: tasks beyond species-level identification and representation learning. 

First, participants agreed on the importance of tasks beyond species identification. These include estimation of species abundance from recordings, automated individual identification, and identification of sounds that have demographic significance and can help inform about population structure and change. 

Second, participants were excited to discuss the use of representation learning, which involves expressing audio as a lower-dimensional embedding that is applicable to many types of sounds. This method holds promise for speeding up classification, improving classification accuracy, and enabling progress for problems with very little training data, like rare species.

Future meetings

AudioXD generated fruitful discussion and collaborations that will benefit the entire bioacoustics community. Many participants were eager to meet again to continue to strengthen connections and speed up progress in our discipline. We hope to have more updates soon on possible future in-person meetings as well as increased virtual coordination. If you’re interested, keep an eye out in this space!

In reply to carlybatist

Are there any papers you/anyone would recommend discussing or utilizing the representation learning you discuss in the 'Frontiers in data analysis methods' section of your summary? I'm keen to look into this more as I haven't really heard much about it yet! 

Maybe not precisely what you're looking for Carly, but Dan Stowell's new paper in PeerJ is a great introduction to computational methods that are being and could be applied to sound classification problems.

Add another post

Want to share your own conservation tech experiences and expertise with our growing global community? Login or register to start posting!