The Joy of BirdNetPi

I don’t think I’ve had as much enjoyment for a piece of software for a very long time as I’ve had with BirdNET-Pi. It’s a realtime acoustic bird classification system for the Raspberry Pi. It listens through a microphone you place somewhere near where you can hear birds, and it’ll go off and guess what it’s hearing, using a cut-down version of the BirdNET Sound ID model. It does this 24/7, and saves the samples it hears. You can then go to a web page (running on the same Raspberry Pi) and look up all the species it has heard.

Our Garden

a somewhat overgrown garden with budding green trees against blue sky

Not very impressive, kind of overgrown, in the wrong part of town. Small, too. But birds love it. At this time of year, it’s alive with birds. You can’t make them out, but there’s a pair of Rose-breasted Grosbeaks happily snacking near the top of the big tree. There are conifers next door too, so we get birds we wouldn’t expect.

We are next to two busy subway/train stations, and in between two schools. There’s a busy intersection nearby, too. Consequently, the background noise is horrendous

What I used

This was literally “stuff I had lying around”:

  • Raspberry Pi 3B+ (with power supply, case, thermostatic fan and SD card)
  • USB extension cable (this, apparently, is quite important to isolate the USB audio device from electrical noise)
  • Horrible cheap USB sound card: I paid about $2 for a “3d sound” thing about a decade ago. It records in mono. It works. My one is wrapped in electrical tape as the case keeps threatening to all off, plus it has a hugely bright flashing LED the is annoying.
  • Desktop mic (circa 2002): before video became a thing, PCs had conferencing microphones. I think I got this one free with a PC over 20 years ago. It’s entirely unremarkable and is not an audiophile device. I stuck it out a back window and used a strip of gaffer tape to stop bugs getting in. It’s not waterproof, but it didn’t rain the whole week it was out the window.
  • Raspberry Pi OS Lite 64-bit. Yes, it has to be 64 bit.
  • BirdNET-Pi installation on top.

I spent very little time optimizing this. I had to fiddle with microphone gain slightly. That’s all.

What I heard

To the best of my knowledge, I have actual observations of 30 species, observed between May 7th and May 16th 2023:

American Goldfinch, American Robin, Baltimore Oriole, Blue Jay, Cedar Waxwing, Chimney Swift, Clay-colored Sparrow, Common Grackle, Common Raven, Gray Catbird, Hermit Thrush, House Finch, House Sparrow, Killdeer, Mourning Dove, Nashville Warbler, Northern Cardinal, Northern Parula, Orchard Oriole, Ovenbird, Red-winged Blackbird, Ring-billed Gull, Rose-breasted Grosbeak, Ruby-crowned Kinglet, Song Sparrow, Veery, Warbling Vireo, White-throated Sparrow, White-winged Crossbill, Wood Thrush

I’ll put the recordings at the end of this post. Note, though, they’re noisy: Cornell Lab quality they ain’t.

What I learned

This is the first time that I’ve let an “AI” classifier model run with no intervention. If it flags some false positives, then it’s pretty low-stakes when it’s wrong. And how wrong did it get some things!

allegedly a Barred Owl, this is clearly a two-stroke leafblower
Black-Billed Cuckoo? How about kids playing in the school yard?
Emergency vehicles are Common Loons now, according to BirdNetPi
Police cars at 2:24 am are Eastern Screech-Owls. I wonder if we could use this classifier to detect over-policed, under-served neighbourhoods?
Great Black-backed Gulls, or kids playing? The latter
Turkey Vulture? How about a very farty two-stroke engine in a bicycle frame driving past?
(This thing stinks out the street, blecch)

There are also false positive for Trumpeter Swans (local dog) and Tundra Swans (kids playing). These samples had recognizable voices, so I didn’t include them here.

The 30 positive species identifications

Many of these have a fairly loud click at the start of the sample, so mind your ears.

American Goldfinch

American Robin

Baltimore Oriole

(I dunno what’s going on here; the next sample’s much more representative)

Blue Jay

Cedar Waxwing

Chimney Swift

Clay-colored Sparrow

Common Grackle

Common Raven

Gray Catbird

Hermit Thrush

House Finch

House Sparrow


Mourning Dove

Nashville Warbler

Northern Cardinal

Hey, we’ve got both of the repetitive songs that these little doozers chirp out all day. Song 1:

and song 2 …

Northern Parula

Orchard Oriole


Red-winged Blackbird

Ring-billed Gull

Rose-breasted Grosbeak

Ruby-crowned Kinglet

Song Sparrow


Warbling Vireo

White-throated Sparrow

White-winged Crossbill

Wood Thrush

Boring technical bit

BirdNetPi doesn’t create combined spectrograms with audio as a single video file. What it does do is create an mp3 plus a PNG of the spectrogram. ffmpeg can make a nice not-too-large webm video for sharing:

ffmpeg -loop 1 -y -i 'birb.mp3.png' -i 'birb.mp3' -ac 1 -crf 48 -vf scale=720:-2 -shortest 'birb.webm'

An hour of Pink Noise

cover made by netpbm, of course
an hour of soothing 2-channel noise

Direct download: 01-pink_noise.mp3

There are a million variations on the simple “use sox to play masking pink noise“, such as:

play -n synth pinknoise gain -3

This will play synthesized pink noise until you hit Ctrl-C.

But it you want two independent noise channels rather than mono, that’s a little more complex. It’s probably easier to download/play the MP3 file above than show you the command line.

Note that MP3s really aren’t designed to encode such random data, and it’s likely that your player will cause the audio to clip in a couple of places. I’m not quite sure why it does this, but it does it repeatably.

If you want to create this for yourself (and create a bonus lossless FLAC, which was far too large to upload here), here’s what I did to make this:



# make the track
sox --combine merge "|sox --norm=-3 -c 1 -b 16 -r 44100 -n -p synth $duration pinknoise" "|sox --norm=-3 -c 1 -b 16 -r 44100 -n -p synth $duration pinknoise" -c 2 -b 16 -r 44100 $outfile fade $fade fade 0 $duration $fade gain -n -3

# make the cover
# 1 - text - 500 x 500 px
pnmcat -white -tb <(pbmmake -white 500 114) <(pbmtextps -font HelveticaBold -fontsize 64 -resolution 180 "PINK" | pnmcrop) <(pbmmake -white 32 32) <(pbmtextps -font HelveticaBold -fontsize 64 -resolution 180 "NOISE" | pnmcrop) <(pbmmake -white 500 114) > cover-text.pbm
# 2 - make the noise bg
pgmnoise 500 500 > cover-noise.pgm
# 3 - make the magenta text
ppmchange black magenta cover-text.pbm > cover-text-magenta.ppm
# 4 - overlay with transparency
pnmcomp -alpha=<(pnminvert cover-text.pbm | pbmtopgm 35 35 ) cover-text-magenta.ppm cover-noise.pgm | cjpeg -qual 50 -opt -baseline -dct float > cover.jpg
# delete the temporary image files, leaving cover.jpg
rm -f cover-text.pbm cover-noise.pgm cover-text-magenta.ppm

# make the mp3
lame -V 2 --noreplaygain -m s --tt 'Pink Noise' --ta 'Pink Noise' --tl 'Pink Noise' --ty $(date +%Y) --tc "scruss, 2021-05" --tn 1/1 --tg Ambient --ti cover.jpg "$outfile" 01-pink_noise.mp3

# make the flac (and delete wav file)
flac --best --output-name=01-pink_noise.flac --delete-input-file --picture=cover.jpg --tag="TITLE=Pink Noise" --tag="ARTIST=Pink Noise" --tag="ALBUM=Pink Noise" --tag="DATE=$(date +%Y)" --tag="COMMENT=scruss, 2021-05" --tag="GENRE=Ambient" --tag="TRACKNUMBER=1" --tag="TRACKTOTAL=1" "$outfile"

You’ll likely need these packages installed:

sudo apt install sox libsox-fmt-all ghostscript gsfonts-x11 netpbm lame flac libjpeg-progs

Mockingbird in the Rain

Northern Mockingbird (Mimus polyglottos) recorded at Centennial College Ashtonbee campus parking lot near Wexford Woods, 07:52 2020-04-13 (rain, handheld phone, noise filtered)

I went out for a very soggy bike ride this morning just to get out of the house. There were a few more people out than I expected, as it’s a regular work day for most people in Ontario. COVID-19 meant that most workplaces were shuttered.

Splashing through the puddles at Centennial College’s deserted Ashtonbee campus (round about here, if you need a precise location) I heard this mockingbird giving its very best performance. I only got a little over a minute of it, but in that time there was some American Robin, Gull, hawk of some kind and best of all (starting just after 40 s) car alarm.

Centennial’s got a big automotive section, and the empty parking lot’s usually full of cars. Mimus was just repeating what it usually heard. I wonder how long they’ll remember and replay car alarms after we’re gone?

A Murder of Crows on your Raspberry Pi with Boodler

Boodler is rather fun. It generates ambient music based on user-defined or downloaded ‘soundscapes’. If you’ve got a modern (HTML5/Opus-capable) browser, you can hear a streaming demo here: It’s using the FM3 Buddha Machine samples in this demo, but it can run lots more: a tree full of crows, a thunderstorm, dripping water, …

It’s pretty easy to run on a Raspberry Pi running a recent version of Raspbian. The only technical glitch I had was that there’s something deeply confused about ALSA sound handling on the Raspberry Pi. I’m sure it’ll get fixed soon, but for now, you have to use PulseAudio. (If you want to read about my ALSA woes, go here.)

The installation prerequisites are simple:

sudo apt-get install pulseaudio pulseaudio-utils libpulse-dev python-dev

Now download and configure Boodler:

tar xvzf Boodler-2.0.4.tar.gz
cd Boodler-2.0.4
python build

It takes a while to do this, but make sure it does something useful when it’s building the various sound drivers. You don’t want it to say:

skipping 'boodle.cboodle_pulse' extension

If it says that, you haven’t installed Pulseaudio. Go back and check your apt-get line.

Once it’s built, now install it:

sudo python install

Now test it:

boodler --hardware --output pulse --testsound

Not merely should you get some pleasant tones from your Raspberry Pi’s audio, but you sound get some informative and non-threatening terminal output. Mine looks like:

Boodler: PulseAudio sound driver.
 PulseAudio library: 2.0.0.
 Sample rate is 44100 fps.
 Samples are 16-bit little-endian.
 Buffer size is 32768.
 21:37:46 (root) Running "Boodler test sound"

If that works, let’s get those crows a-cawin’. Download the soundscapes you need:

boodle-mgr install
boodle-mgr install

and run it:

boodler --output pulse com.eblong.zarf.crows/ParliamentOfCrows

Crows everywhere!

I really like the Buddha Machine samples. It’s quite big (> 80 MB), so this next set will take a while to download:

boodle-mgr install
boodle-mgr install

It’s worth the wait:

boodler --output pulse com.azulebanana.buddhaagent/ChangingLoops

Boodler has tons of options, prebuilt packages, and instructions to build your own: Boodler Documentation.

One thing I’ve tried to get working, but failed, is streaming from Boodler via icecast. Sure, I can install and run it, it’s just that the results are, um, undesirable. If you want to have a play, here’s how to install icecast:

sudo apt-get install icecast2 ices2 libshout3-dev

Icecast will configure itself, and ask for a couple of passwords. You’ll have to rebuild and reinstall Boodler for it to catch the new configuration. You can then try streaming:

boodler --output shout --define shout-password=mypassword --define shout-mount='/boodler-buddha.ogg' com.azulebanana.buddhaagent/ChangingLoops

If you open a web browser at this address http://raspberrypi:8000/ you should see a config page listing your boodler-buddha.ogg stream. Click on the M3U link next to it, and your streaming music player should start making a joyful noise …

… except in my case, something went very wrong, and it started to produce industrial ultra-glitch nightmare noise: boodler-streaming_test-fail. I’m sure it’s fixable with some tweaking, but I’m not there yet.