Now that Google Summer of Code has come to an end, I would like to take a moment to share my experience over the last three months. I learned a lot during this time from my mentors as well as from my fellow contributors.
I am glad I decided to contribute to Orcasound and met such a wonderful community. Orcasound is an amazing community consisting of experts and learners from various professions and walks of life. The mentors are extremely helpful and friendly. The community on the Slack channel is also quite active.
We also met many amazing people outside Orcasound. For example, Scott introduced us to another team called HALLO based in Canada which was also working on similar projects. They were working on open labeled data and models related to orcas and humpbacks at the time. They invited us to attend their bi-weekly calls to introduce ourselves and discuss our projects. The HALLO workshops were fun and it was nice meeting the Canadian team as we tried to find some commonalities in our works.
We also had another meeting where Valentina introduced us to Tom Denton from Google whom she met during a workshop. Tom has been working on bioacoustics for quite some time. The meeting was very informative and he even suggested to me a few models as he had previously worked on source separation for birds.
It was overall an extremely delightful experience. But that doesn’t mean there were no roadblocks. As mentioned in my previous blog post, the biggest roadblock in the pathway to success for this project was the unavailability of isolated orca vocals that were required to train the neural networks. Hence, most of my time was spent creating a nice balanced dataset. I used open source tools such as Audacity to extract orca vocals from hydrophone recordings. I even used some of Ambra’s de-noising techniques to extract orca vocals. In the initial stages, I had to manually go through a lot of hydrophone recordings to put together a set of high-quality SRKW calls with little background noise. You can go through my HALLO workshop presentation slides and the wiki page of the acoustic-separation repository for more information about the dataset and models used.