Intro to my project, then how to improve predictions for morning/evening choruses?

[BradGilliam]

Hello,

Here is an intro to my situation: I have a ~75 TB dataset of acoustic recordings (primarily targeting birds) spanning 20 years (and counting), mainly from a wooded property in SE Ohio. I am neither a biologist/scientist, nor am I an ML expert, and nor am I a Python programmer, and I haven’t used Raven Pro (yet). My background is electrical/electronics and systems engineering, and I’ve attended two Recording Workshops held by the Macaulay Library (2005, 2007). My initial take is that the BirdNET Analyzer may be just the ticket for processing my large dataset.

I have installed the new Windows GUI to start with. The GUI seems to be providing decent predictions from my test tracks. CSO parameter settings at 0.9, 0.5, 0.0 seem to suffice, although I’ll continue to evaluate these before settling on a set and launching into the full-scale dataset analysis effort. Additionally, I have the GUI set to use a custom species list (the 2022 ABA checklist converted to the .txt format required by the Analyzer), and the output is exported via CSV files for which I have created VBA macros to organize and make clearer sense of the prediction results.

Key observation: For a given mid-May morning chorus track recorded in SE Ohio (~5-6 AM), I’m noticing that the Analyzer returns only one species prediction for a given 3-second interval, when in fact there are multiple species vocalizing simultaneously within close proximity to the mics during these 3-second intervals. It appears that the loudest bird is the one that gets predicted.

I’ve reviewed the README & Guide but I could still use some expert advice:

1. Is it possible to have the Analyzer return predictions for say, 3 or 4 of the species that are vocalizing simultaneously within a given 3-second interval? This is something that I highly desire, especially for morning and evening choruses. I anticipate that this might require a move to a CLI form of the Analyzer? Seeing that there is available birdnet, birdnetlib, and birdnetR, I’m not sure which way to go. What recommendations would you experts have to steer me in the right direction?
2. Nearly all of my recorded tracks are 24-bit 96 kHz L&R channels combined and saved as .WAV files by the recorder, are there any problems to be expected for the Analyzer with this configuration?

For some recording seasons where I implemented shotgun mics pointed in separate directions, I’ll need to use WaveLab Elements to separate out those L&R channels, creating ever more tracks within the dataset for analysis.

(Minor thought: has there been any contemplation regarding having the CLO host a hands-on BirdNET workshop, assuming you haven’t yet held one?)

Thanks in advance for any help you may provide!

Brad

[Mattk70]

Hey @BradGilliam , regarding multiple species detections, you should read this thread: #541.

For such a large dataset, you may want to consider software which runs BirdNET inference more rapidly than the Analyzer GUI, such as BirdNET Go, or Chirpity.