I'm working on creating a gunshot classifier that is (hopefully) able to distinguish between three different gun types that are used for different poaching purposes in Central Africa (large game vs. small game). I've collected recordings in Central Africa of gunshots originating from these three gun types at different distances and orientations from the gunshot origin, in different vegetation types, and at various temperature and humidity levels to encompass as much variation as possible in how the recorded gunshots for each gun type may vary based on these factors. I am currently at the data processing step but will next need to determine what acoustic parameters might be the most useful to extract from the gunshots (which are broadband signals). My previous bioacoustics work has not focused on analyzing broadband signals so I am interested to see if others have worked with these types of signals before and whether folks may have insight into useful features I can focus on extracting for the machine learning model I am hoping to develop.
I have not yet tried to extract any acoustic parameters from the gunshot signals and am hoping to gain some insight into what features might be the most useful to focus on prior to taking the next steps in my research.
Some authors (including Freire and Apolinario 2010 and Hrabina and Sigmund) find mel-frequency cepstral coefficients to be particularly helpful for gunshot detection & classification.
I analyze broadband sounds of underwater explosions as well as broadband echolocation clicks. I would recommend calculating the following features: Peak frequency, center frequency, bandwidth at 3 dB, bandwidth at 10 dB, and duration.
PAMpal is an R package which can help you calculate features commonly used in echolocation click analysis. https://github.com/TaikiSan21/PAMpal
These types of questions might be relevant for the proposed Bioacoustics Stack Exchange proposal. Any other readers are invited to join us there!