RTSP noise/buffering issue

[stevedee]

I have a rtsp stream from my BirdNet-Pi system.
If I listen using my Linux laptop or my phone, the sound quality is good.
But with my 2 BirdNet-Go systems I get a crackling noise: https://icedrive.net/s/C58yivYRw5avXGwPFxx1RTxP5567

I wonder if this is a buffering issue on BirdNet-Go; is buffering used or can I add/adjust the size some how?

[stevedee]

As a double check I ran this command on BirdNet-Go:-
ffmpeg -rtsp_transport tcp -i rtsp://192.168.1.149:8554/mic -t 30 out.wav

...and the out.wav file is clean, no clicks!

[stevedee]

...also tried pre-processing & buffering like this:-

ffmpeg -rtsp_transport tcp -i rtsp://192.168.1.149:8554/mic -rtbufsize 200M -ar 44100 -ac 1 -f wav pipe:1 | /home/steve/birdnet-go-app/birdnet-go realtime -

Still clicks on Birdnet-Go

[tphakala]

Can you please be more specific, is clicking impacting recordings, or live playback, or both?

[stevedee]

Its certainly affecting Live Playback, as you can hear from the recording (of live playback) in post#1.

I assume its affecting detections, as I'm currently comparing 2 x BirdNet-Go with a BirdNet-Pi system and getting different results between all systems; this is why I started listening to the audio to see if I had a streaming problem (although not a streaming problem with laptop, phone, or even on one of the Birdnet-Go Pi computers when using ffmpeg).

[tphakala]

Okay, this is somewhat known issue, live playback has had quite lot of issues and these audio glitches should not happen for recordings or affect detections, it happens only on live stream level. Are you using old UI in basepath / or are new UI which is available in /ui/, new UI live playback player has many bug fixes.

When it comes to different results between BirdNET-Go and -Pi, they are quite different in way detections are aggregated. For example BirdNET-Go employs deep detection by default which requires multiple matches per species before accepting detection as valid, this is to limit number of false positive detections. If you wish to disable this, set BirdNET overlap value to 1.5 or lower in BirdNET-Go settings, then you will see much more detections but also much more false positives.

[stevedee]

So I've just checked recordings (I should have done this sooner!) and they appear to be 'click' free.
Live playback is much the same on both standard & new UI.

When I started this comparison, I had no idea what I'd find, if anything. That's why I added a 2nd BirdNet-Go, to check not only the variations with Pi but also those between 2 Go systems.
All 3 systems have an overlap = 1.0

Some of the results for this morning (from 5am til 9am) look like this:-

_______Pi - Go#1 - Go#2
Swift; 101 - 44 - 51
Magpie; 32 - 12 - 14
Scoter; 4 - 1 - 1
wood-pigeon; 24 - 18 - 17
Carion Crow; 5 - 5 - 9

The reason for the high 'out of season' swift detections is that we have a new swift box with internal swift caller. However, BNP picks up twice as many of these fake calls as the BNG.

I'm probably wasting my time, but its fun.
I view these systems as sampling devices, so you just try to do your best when it comes to siting the microphone and configuring the software. But maybe the variations between systems (even when sharing the same audio source) are just a fact of life!

[tphakala]

Live playback is much the same on both standard & new UI.

Are you using latest nightly release? I improved live streaming for this release https://github.com/tphakala/birdnet-go/releases/tag/nightly-20250805 - live playback was much improved by this PR #1087 and I don't hear any glitching on from my RPI 5 host on macOS or Windows clients.

If you are not yet familiar of update process you can update simply running install.sh again, it will detect existing deployment and offer update option.

[stevedee]

I'm using nightly-20250813-1-ge680d30f on both systems.
One is an image install, the other docker. Do they both update using the install.sh method?

[tphakala]

install.sh manages docker deployments only