Susurrant (soo-SIR-ant) is a tool for analyzing collections of sounds, especially those that are part of a social network.
Input your audio, text, and social network data, then navigate a machine-generated index of the distribution of "topics" within your corpus.
By sonifying its internal feature representations as well as the outputs from its algorithm, Susurrant aims to give you a better understanding of the machine listening process from start to finish.
Susurrant is and will always be free. The code is released under the Affero GPLv3 License and is available on GitHub.
This software is far from a proper release, but development details (roadmap, etc.) can be found on the main repository's wiki. I welcome input about how this software might be useful to you, and especially welcome contributions of documentation/code!
See Install for (terse) instructions on how to run Susurrant locally, or visit the demo for a taste of how it works (currently without sound).
There’s a fun post over on Google’s Research Blog detailing the “dreams” of neural networks, generated by feeding random noise images into the networks and ‘nudging’ the image iteratively toward a certain classification. While the details are somewhat different, this is a great visual analogue for the processes used in Susurrant to expose the workings of its own audio “interpretation” algorithms.
I’ve just returned from a wonderful conference at UCLA called “Inertia: A Conference on Sound, Media, and the Digital Humanities,” organized by Mike D’Errico and company. On Friday, I presented the notion of “algorithmic listening” (i.e., the practice of listening to algorithms that listen) and its application in the context of my development of Susurrant.