Backyard Bird Identification
There is an AI audio based model that has been developed called BirdNET
BirdNET is a research platform that aims at recognizing birds by sound at scale. We support various hardware and operating systems such as Arduino microcontrollers, the Raspberry Pi, smartphones, web browsers, workstation PCs, and even cloud services. BirdNET is a citizen science platform as well as an analysis software for extremely large collections of audio. BirdNET aims to provide innovative tools for conservationists, biologists, and birders alike. -BirdNET
A project BirdNET-Pi exists that has created an image to run this model and various web services on a raspberry pi.
I already have a server running so I was hoping to run the model in something more portable, ideally a docker image. After a quick search I found BirdCAGE which repackages the components into a set of docker images.
My docker-compose.yml:
services:
birdcage_redis:
container_name: birdcage_redis
image: "redis:latest"
restart: always
birdcage_backend:
container_name: birdcage_backend
image: mmcc73/birdcage_backend:latest
depends_on:
- birdcage_redis
environment:
DATABASE_FILE: /db/birdcage.db
API_SERVER_PORT: 7007
TEMP_DIR_NAME: tmp
ANALYZE_SERVER: 192.168.1.XXX
ANALYZE_PORT: 7667
DETECTION_DIR_NAME: detections
CORS_ORIGINS: http://192.168.1.XXX:7008
REDIS_SERVER: birdcage_redis
REDIS_PORT: 6379
CONCURRENCY: 10
JWT_SECRET_KEY: xxxxx
PULSE_SERVER: 172.17.0.1
tmpfs:
- /tmp:size=16M
volumes:
- ${DOCKER_VOLUMES}/birdcage/detections:/detections
- ${DOCKER_VOLUMES}/birdcage/db:/db
- /etc/localtime:/etc/localtime:ro
- /etc/timezone:/etc/timezone:ro
ports:
- 7007:7007
restart: always
birdcage_frontend:
container_name: birdcage_frontend
image: mmcc73/birdcage_frontend:latest
depends_on:
- birdcage_backend
environment:
API_SERVER_URL: http://192.168.1.XXX:7007
WEBUI_PORT: 7008
volumes:
- /etc/localtime:/etc/localtime:ro
- /etc/timezone:/etc/timezone:ro
ports:
- 7008:7008
restart: always
birdnetserver:
container_name: birdnetserver
image: mmcc73/birdnetserver:latest
ports:
- "7667:8080"
restart: always
Once the containers were running, I was able to access the web-ui and now it was time to figure out an audio source.
I wanted to separate the audio/mic source to be low power and easily movable for positioning. I opted to use an omnidirectional lavalier microphone into an old pi zero w I had lying around.
https://www.amazon.com/dp/B07RMXR4FB
In order to create the RTSP stream from the microphone, I used go2rtc
My go2rtc.yaml config file:
streams:
mic:
- exec:ffmpeg -hide_banner -f alsa -channels 1 -sample_rate 48000 -i hw:0,0 -c:a libopus -rtsp_transport tcp -f rtsp {output}
This creates a single RTSP stream using the microphone via alsa. At first I used AAC but found it a bit too taxing for this single core CPU. Switching to OPUS as the codec lowered the CPU usage quite a bit.
Once I had the RTSP stream up and running I tested it briefly with ffprobe and VLC
❯ ffprobe rtsp://192.168.1.110:8554/mic
ffprobe version 7.0.1 Copyright (c) 2007-2024 the FFmpeg developers
built with Apple clang version 15.0.0 (clang-1500.3.9.4)
configuration: --prefix=/opt/homebrew/Cellar/ffmpeg/7.0.1 --enable-shared --enable-pthreads --enable-version3 --cc=clang --host-cflags= --host-ldflags='-Wl,-ld_classic' --enable-ffplay --enable-gnutls --enable-gpl --enable-libaom --enable-libaribb24 --enable-libbluray --enable-libdav1d --enable-libharfbuzz --enable-libjxl --enable-libmp3lame --enable-libopus --enable-librav1e --enable-librist --enable-librubberband --enable-libsnappy --enable-libsrt --enable-libssh --enable-libsvtav1 --enable-libtesseract --enable-libtheora --enable-libvidstab --enable-libvmaf --enable-libvorbis --enable-libvpx --enable-libwebp --enable-libx264 --enable-libx265 --enable-libxml2 --enable-libxvid --enable-lzma --enable-libfontconfig --enable-libfreetype --enable-frei0r --enable-libass --enable-libopencore-amrnb --enable-libopencore-amrwb --enable-libopenjpeg --enable-libspeex --enable-libsoxr --enable-libzmq --enable-libzimg --disable-libjack --disable-indev=jack --enable-videotoolbox --enable-audiotoolbox --enable-neon
libavutil 59. 8.100 / 59. 8.100
libavcodec 61. 3.100 / 61. 3.100
libavformat 61. 1.100 / 61. 1.100
libavdevice 61. 1.100 / 61. 1.100
libavfilter 10. 1.100 / 10. 1.100
libswscale 8. 1.100 / 8. 1.100
libswresample 5. 1.100 / 5. 1.100
libpostproc 58. 1.100 / 58. 1.100
[rtsp @ 0x15a004e70] method SETUP failed: 461 Unsupported transport
Input #0, rtsp, from 'rtsp://192.168.1.110:8554/mic':
Metadata:
title : go2rtc/1.9.2
Duration: N/A, start: 0.000000, bitrate: N/A
Stream #0:0: Audio: opus, 48000 Hz, stereo, fltp
Once it was confirmed working and I could hear sound, I added the stream as a source to BirdCAGE.
BirdNET has some tunables so I adjusted lat/long which are used as algorithm inputs, along with the accepted confidence level.
After some time running, I started to get detections!
There is a section that shows a detection count split by time of day
Here are the details shown per detection (along with a 6 second recording)
Soon after writing this initial post, I switched to https://github.com/tphakala/birdnet-go and it is a REALLY clean, single binary, nice looking tool for running the same birdNET models!
