Simon Dobson (Posts about project:acoustic-birds)

Processing MicroMoth recordings offline

Simon Dobson — Mon, 10 Jun 2024 11:57:22 GMT

Processing MicroMoth recordings offline

The uMoth generates .wav files, uncompressed waveforms of what it records. These need to be processed to identify any bird calls within them.

This function is integrated in BirdNET-Pi, which does recording and classification, and provides a web GUI. With the uMoths we need to provide classification as part of a data processing pipeline. We can however make direct use of the classifier "brain" within BirdNET-PI, which is unsurprisingly called BirdNET-Analyzer.

Installation

I'm working on a 16-core Intel Core i7@3.8GHz running Arch Linux.

First we clone the BirdNET-Analyzer repo. This takes a long time as it includes the ML models, some of which are 40MB or more.

    git clone https://github.com/kahst/BirdNET-Analyzer.git
    cd BirdNET-Analyzer

The repo includes a Docker file that we can use to build the analyser in a container.

    docker build .

The container setup is quite basic and is probably intended for testing rather than production, but it gives a usable system that could then be embedded into something more usable. The core of the system is the analyze.py script.

Analysing some data (AKA identifying some birds!)

The container as defined looks into its /example directory for waveforms and analyses them, generating text file for each sample. The easiest way to get it to analyse captured data is to mount a data directory of files onto this mount point (thereby shadowing the example waveform provided).

There are various parameters that configure the classifier. I copied the defaults I was using with BirdNET-Pi, only accepting classifications at or above 0.7 confidence.

    docker run -v /var/run/media/sd80/DATA:/example birdnet-analyzer analyze.py --rtype=csv --min_conf=0.7 --sensitivity=1.25

This crunches through all the files (982 of them from my first run) and generates a CSV file for each. An example is:

Start (s)	End (s)	Scientific name	Common name	Confidence
6.0	9.0	Corvus monedula	Eurasian Jackdaw	0.9360
9.0	12.0	Corvus monedula	Eurasian Jackdaw	0.8472
12.0	15.0	Corvus monedula	Eurasian Jackdaw	0.8681
15.0	18.0	Corvus monedula	Eurasian Jackdaw	0.8677
24.0	27.0	Columba palumbus	Common Wood-Pigeon	0.9198
27.0	30.0	Columba palumbus	Common Wood-Pigeon	0.7716
45.0	48.0	Corvus monedula	Eurasian Jackdaw	0.8023
48.0	51.0	Corvus monedula	Eurasian Jackdaw	0.7696

Those are entirely credible identifications. The start- and end-point offsets allow rough location within the recording. (BirdNET segments the recordings into 3s chunks for analysis.)

This is clearly not as straightforward as BirdNET-Pi, nor as immediately satisfying. But it does scale to analysing lots of data (and could be made to do so even better, with a better front-end to the container), which is important for any large-scale deployment.

Deploying a MicroMoth

Simon Dobson — Mon, 10 Jun 2024 10:11:02 GMT

Deploying a MicroMoth

The MicroMoth (or uMoth) from is the same as their better-known AudioMoth recorder but with a significantly smaller footprint. It's just a traditional recorder or data-logger, with now on-board analysis and no wireless connectivity. I got hold of some to use in a larger project we're thinking about running, and they're not kidding about the "micro" part.

The uMoth uses the same software as the AudioMoth, and therefore the same configuration app available from the apps page – for 64-bit Linux in my case. It downloads as a .appimage file, which is simply a self-contained archive. It needed to be marked as executable, and then ran directly from a double-click. (The page suggests that there may be some extra steps for some Linux distros: there weren't for Arch.)

I then followed the configuration guide. The time is set automatically from the computer's clock when you configure the device.

For testing I chose two recording periods, 0400–0800 and 1400–1600.

As shown this will, with the default 48KHz sampling, generate about 2GB of data per day and use about 70mAh of energy. For my tests I just hung the device out of the window on a USB tether for power: it works fine drawing power from the USB rather than from the battery connector.

This turned out not to record anything, because the time is lost if the power is disconnected, even though the configuration is retained. (The manual does actually say this, with a suitably close reading. It could be clearer.) There's a smartphone app that can reset the time once the device is in the field and powered-up, though, by making an audio chime that encodes the current time and location in a way the board can understand. Flashing the device with the "Always require acoustic chime on switching to CUSTOM" makes it wait after power is applied until its time is set.

The red LED flashes when the device is recording. The green LED flashes when the device is waiting for a recording period to start. The red LED stays lit while the time is unset.

First installation of BirdNET-Pi

Simon Dobson — Sun, 19 May 2024 14:19:45 GMT

First installation of BirdNET-Pi

The BirdNET-Pi system aims to provide out-of-the-box bird identification. It's slightly more awkward than that, but still pretty straightforward to get up and running.

My first hardware plan was to use a Raspberry Pi Zero as the compute host with a Waveshare WM8960 HAT for the sound capture. It turns out that BirdNET needs a 64-bit platform – why I'm not sure – and the Pi Zero only runs 32-bit Linux. I therefore moved to a Raspberry Pi B that I had lying around, and put a 64-bit "lite" install on it to run headless.

I then basically just followed the installation guide. There was an issue with the installation script when cloning the GitHub repo: I suspect this was because of limited memory on the Pi. I downloaded manually, and manually ran the rest of the install script, which did a lot of setup of services and a PHP web server.

I compiled the drivers for the HAT, which worked fine. The new sound card is recognised but is not the system default.

The installed components seem to include:

icecast2, a streaming server, used to replay recordings
caddy web server
PHP for serving the web pages
arecord to actually record audio
ffmpeg to extract waveforms
sqlite for the database
the actual machine learning model used for recognition

The recognition models are built with TensorFlow. This is a great example of how the standard Linux tools and services can be combined to get a scientific-grade sensor platform. (Caddy doesn't seem to be running over TLS by default, which would be an issue outside a firewall.)

Since the sound card isn't the default, the easiest way to get the system listening to the right mics is to change the device in the "advanced settings" panel: in my case I changed from "default" to "hw:2,0", reflecting the output of arecord -l that shows the sound card devices.

I then deployed the Pi out of the kitchen window.

To start with it wasn't hearing anything, which I think may be because of the waterfall in the courtyard: turning this off made things much more effective:

That's an appropriate set of birds being seen – and we hardly ever see magpies, but know they're around. There's actually quite a lot of background noise even in such a quiet village, but the bird calls do stand out.

I can't see any reason for the manual installation on bare metal: as far as I can see everything could be containerised, which would make deployment and management a lot easier.