The people around Open Source Beehives (OSBH) were directing towards audio from the very beginning and therefore presenting coefficients which where won via machine learning, learned via the audio-samples they had so far.
The output is simply the activity/health status of a bee colony. So far the algorithm can tell whether the colony is dormant, active, pre-, post- or swarming, if the queen is missing or hatching. For more background information about the audio processing, please follow up reading current work status thread in the OSBH Forum. So far, the results are promising.
We forked the "OSBH machine learning" repository to osbh-audioanalyzer to make it able to obtain an input file option.
For more information, see also Rate vitality of bee colony via analysing its sound.
audiohealth analyze --audiofile ~/audio/samples/colony_before_swarming_25_to_15.ogg --analyzer tools/osbh-audioanalyzer/bin/test
Output:
==================
Sequence of states
==================
pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, active, active, active, active, active, active, active, active, active, active, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, pre-swarm, active, active, active, active, active, active, active, pre-swarm, active, pre-swarm, pre-swarm, pre-swarm, pre-swarm, active, active, active, active, pre-swarm, pre-swarm, active, active, pre-swarm, active, pre-swarm, active, active, active, pre-swarm,
===================
Compressed timeline
===================
0s - 80s pre-swarm ========
90s - 180s active =========
190s - 310s pre-swarm ============
320s - 380s active ======
390s - 400s pre-swarm =
400s - 410s active =
410s - 440s pre-swarm ===
450s - 480s active ===
490s - 500s pre-swarm =
510s - 520s active =
530s - 540s pre-swarm =
540s - 550s active =
550s - 560s pre-swarm =
560s - 580s active ==
590s - 600s pre-swarm =
==============
Total duration
==============
320s pre-swarm ================================
280s active ============================
======
Result
======
The most common events (i.e. the events with the highest total duration) are:
The colony is mostly in »PRE-SWARM« state, which is going on for 320 seconds.
Sometimes, the state oscillates to »ACTIVE«, for 280 seconds in total.
$ audiohealth --help Usage: audiohealth analyze --audiofile audiofile --analyzer /path/to/osbh-audioanalyzer [--strategy lr-2.1] [--debug] [--keep] audiohealth analyze --wavfile wavfile --analyzer /path/to/osbh-audioanalyzer [--strategy lr-2.1] [--debug] audiohealth analyze --datfile datfile --analyzer /path/to/osbh-audioanalyzer [--strategy lr-2.1] [--debug] audiohealth convert --audiofile audiofile --wavfile wavfile audiohealth spectrogram --audiofile audiofile --pngfile pngfile audiohealth power --audiofile audiofile --pngfile pngfile audiohealth power --wavfile wavfile --pngfile pngfile audiohealth --version audiohealth (-h | --help) Options: --wavfile=<wavfile> Name of .wav file --pngfile=<pngfile> Output .png file of power spectrum --audiofile=<audiofile> Process audiofile. Please use sox-compatible input formats. --datfile=<datfile> Process datfile. --analyzer=<analyzer> Path to OSBH audioanalyzer binary --strategy=<strategy> The classification strategy. One of dt-0.9, dt-1.0, dt-2.0, lr-2.0, lr-2.1 --keep Keep (don't delete) downsampled and .dat file --debug Enable debug messages -h --help Show this screen
Hint: By using --strategy dt-2.0 or even --strategy dt-1.0, different
classification strategies can be toggled to be able to compare results against each other.
"dt-" means "decision tree", while "lr-" means "logistic regression".
The cutting edge classification strategy is "lr-2.0", which is also the default setting.
See also commit Added new filters and logistic regression classifier. Aaron Makaruk describes it on 2017-07-01 like:
Our classifier has been recently updated to include two new states, and we've moved past decision-tree algorithms to something yielding greater results.
git clone --recursive https://github.com/hiveeyes/audiohealth cd audiohealth
To relieve your machine from compiling SciPy or NumPy, install the python libraries from your distribution. audiohealth furthermore relies on sox for audio resampling. We also recommend youtube-dl for downloading audio samples from Youtube.
Install some distribution software packages (if you are not root, you might want to prefix the command with sudo):
apt install python-virtualenv python-scipy python-numpy sox libsox-fmt-all xvfb xauth youtube-dl
Build the osbh-audioanalyzer:
make setup-osbh-audio-analyzer
make setup-virtualenv source .venv/bin/activate
# Acquire example files. wget https://community.hiveeyes.org/uploads/default/original/1X/6de56aed3b520166d92ce3194ff6a9e6852491b2.mp3 -O colony-with-queen-gruber.mp3 wget https://community.hiveeyes.org/uploads/default/original/1X/c751a4a83e8c1f5f0380eee2c7ee310b0967056a.mp3 -O colony-without-queen-gruber.mp3 # Convert to .wav files. sox colony-with-queen-gruber.mp3 colony-with-queen-gruber.wav sox colony-without-queen-gruber.mp3 colony-without-queen-gruber.wav # Run OSBH audio analyzer audiohealth analyze --audiofile=colony-with-queen-gruber.mp3 --analyzer=tools/osbh-audioanalyzer/bin/test audiohealth analyze --audiofile=colony-without-queen-gruber.mp3 --analyzer=tools/osbh-audioanalyzer/bin/test # Compute power spectrum and find peaks audiohealth power --audiofile=colony-with-queen-gruber.wav --pngfile=power.png open power.png # Compute spectrogram audiohealth spectrogram --audiofile=colony-with-queen-gruber.wav --pngfile=spectrogram.png open spectrogram.png
The driving force behind the audio signal processing at OSBH is Javier Andrés Calvo, so we want to send a big thank you to him and the whole OSBH team - this program is really standing on the shoulders of giants. Keep up the good work!