extract-positives: a handy tool for pulling detections from recordings

bioacoustics

Published

July 20, 2026

Modified

July 20, 2026

TL;DR

extract-positives is a python utility that you can use to extract the source audio corresponding to positive frames in buzzdetect results. Extract a whole dir, filter by time, add additional context on either side of the frames, and more!

When we run buzzdetect analyses on large datasets, we usually wrangle the results with buzzr and do some sanity checks, but we let buzzdetect do the listening and don’t check the raw audio ourselves. Occasionally, we want to take a closer listen. Sometimes this is because we want to get a clue as to which insects are making the buzzes. Other times, it’s because we’re suspicious of the results. For example, in developing a new model (not the currently deployed model_general_v3) for improved bumble bee detection, I was noticing extraordinary spikes in buzz detections right at 6:00 PM in every single audio file from a deployment. This was, let’s say, entomologically implausible. My erstwhile method for investigating an issue like this was to open the problematic files one at a time, seek to the approximately correct position, listen for a while, and write down my notes on a scratch pad. That was so 2025. To make my life a little easier, I built the extract-positives utility.

This audio is from some unpublished data I’m using in model development. Can you spot the difference? Left of the playhead are true buzzes. To my ear, the first three are bumble bees and the next four are honey bees. The next, much louder frames were the source of our mysterious false positive spikes…the home owner doing some weed whacking! You can see why a bioacoustic model might confuse these, at a glance. Note that the second and fourth honey bee segments are longer than the others; that’s because there were contiguous detections that extract-positives merged together.

A demo: five-flowers

Let’s use the Five Flowers results and audio for a little demo. I’m going to put the audio next to the snip_results and rename to snip_audio, but it doesn’t really matter where the folders are.

Now let’s launch the current version of extract-positives at the time of writing:

Okay, it’s ugly as sin but it does run. Some notes:

Ensure Audio is set to Folder and point it at your audio dir
Ensure Results is set to Folder and point it at your results dir
Set Join output to “By recorder” and point your output folder wherever you like.
- “By file” means that one input file produces one output file of extracted positives.
- “By recorder” means that each input recorder (in our file storage schema, the dir one level up from the recordings is always named with the ID of the recorder that made the recordings)
- “Join all” concatenates all frames into one big file
For this run, we won’t do anything with time. Leave the time filter columns blank.
I like outputting to MP3 since lossy compression isn’t a big concern for spot-checking results and the files can become quite large.
Set your threshold to -1.2, since this is our 95% precise threshold for model_general_v3

Full-length audio files could carry thousands of positive frames, so it’s best to leave that “Check output size” box checked.

Hit the check button and the tool will count the number of detections and estimate the output size:

Once you hit run, the files will be written to your chosen dir. Let’s check out

Opening the file in SeeNote gives:

Nice! Couldn’t be easier. Some notes:

The black bars are from that dead time setting, which lets us differentiate frames
Wider segments are from contiguous positive frames; the tool merges them together, because otherwise the buffer time on either side would cause extracted frames to overlap and repeat

Happy listening!

TL;DR

A demo: five-flowers

Coda: what’s the value of sharing, anymore?