A queenless colony has weeks to live, and beekeepers have long said they can hear it. Published classifiers agree, with accuracies in the high nineties. This repository asks a narrower question. Someone who wants to put a microphone in their hive needs to know whether any of that survives a colony the model has never heard.
Three readings of the same recordings go through the same three tests. The first is a log-mel convolutional network, which is what the literature uses. The second is eight band energies, which is what a beekeeper's chart shows. The third is a physical model of the hum built from wing beats. That model can only explain a spectrum through the harmonics of a wing stroke. It returns per-clip quantities that have a physical meaning: the colony's wing-beat frequency, its spread, and the share of the sound that is bees at all.
Hive acoustics came to me from the field: beekeepers and agricultural clients who lose colonies because nobody notices a queenless hive until it is too late. A hive tells you everything through its hum, if you know which physics to listen for. I have worked on that signal for years in research and client projects, and this repository puts the core of it in the open: a wing-beat model of the colony built from first principles and tested on public recordings from six hives.
Inês Nolasco and Emmanouil Benetos, To bee or not to bee: an annotated dataset for beehive sound recognition (Zenodo 1321278, CC BY 4.0). There are two sources, and the difference between them is the whole design:
| source | hives | sessions | what a label is |
|---|---|---|---|
| NU-Hive | 2 (Hive1, Hive3) | 5 | the same hive recorded on separate days with the queen present and after she was removed |
| Open Source Beehives | 4 citizen-science hives | 4 | one hive, one state: Active or Missing Queen |
Every file carries an annotation of which seconds are hive sound and which are a passing car or a voice. Only the hive-sound seconds are used, cut into two-second clips at 8 kHz. That gives 56 files, 10 277 clips and 5.7 hours. After capping each session at 600 clips, so that no single afternoon dominates, 4 301 clips remain.
Nine sessions, six hives. That is the true sample size for any claim about queens, and every number below should be read with it in mind.
| # | Finding | Evidence |
|---|---|---|
| 1 | Two thirds to four fifths of a hive's acoustic energy is below 260 Hz, and the sessions of one class differ from each other by more than the classes differ. The textbook 180-260 Hz "worker flight" band holds 7 % of the power in queenright and 6 % in queenless sessions. | section 3 |
| 2 | A random split over clips gives AUC 0.9999. The same network, asked about a session it has not heard, gives 0.05; about a hive it has not heard, 0.38. Below chance is not noise. The network learned to recognise sessions: no held-out session is called right by a majority of its clips, and every queenless session scores above every queenright one (2 of 126 relabellings are that extreme, p = 0.016). The "nearest session has the other label" explanation given first covers five of the nine sessions at most. | sections 4, 4.1 |
| 3 | The hum is well described as a harmonic comb - a learned single-bee spectrum with a 2.4 Hz line width that widens with harmonic number. But the model places most of the colony's fundamental at 100-160 Hz, not at the textbook 200-250 Hz - in the five NU-Hive sessions. In the four citizen-science hives it piles the mass against the 80 Hz floor of its grid instead. Either these recordings hum lower than the literature says, or the model is reading something periodic that is not a bee (see below). | sections 5, 5.1 |
| 4 | Five named quantities from the wing-beat model transfer where the network does not: AUC 0.60 across sessions and 0.62 across hives, against 0.52 / 0.57 for band energies. That is above chance and nowhere near a product. | section 5 |
| 5 | The full wing-beat distribution p(f0), fed to the same regression, does worse than chance across sessions (0.20). Fifty-five numbers per clip are enough to learn session identity again. The physics helps only when it is reduced to the few quantities that mean something. | section 5 |
One curve per recording session, area-normalised, blue queenright and orange queenless. Two things are visible before any model is fitted. The energy is low: most of it sits under 260 Hz. The shape of the spectrum differs between the two projects (different microphones, boxes and countries) more than between the two queen states. And within a project, the two queenless sessions of Hive3 do not look more like each other than either looks like the queenright session.
These are the bands the literature names, one line per session. The thing being replicated is the session, not the clip. Ten thousand two-second clips from one afternoon in one hive are one observation of one colony. Treating them as independent is how a difference between two recordings becomes a p-value with ten zeros in it. Read that way, no band separates the classes by more than the spread within a class. The 500-600 Hz "swarming band" holds 2.7 % versus 1.3 % - a difference, with a standard deviation across sessions of 1.5 and 1.0.
Because there are only nine sessions, "larger than the spread" can be made
exact. Four queenright labels among nine sessions can be handed out in
C(9, 4) = 126 ways. So every relabelling is enumerated and the observed gap
between class means is placed among all 126 (src/band_permutation.py,
results/band_permutation.json). The smallest two-sided p-value nine
sessions can produce is 1/126 = 0.008. No band comes close:
| band | queenright | queenless | session AUC | exact p | p without CF001 |
|---|---|---|---|---|---|
| 20-100 Hz | 0.33 | 0.43 | 0.45 | 0.61 | 0.94 |
| 100-180 Hz | 0.26 | 0.30 | 0.45 | 0.71 | 0.57 |
| 180-260 Hz (worker flight) | 0.069 | 0.061 | 0.65 | 0.76 | 0.91 |
| 260-400 Hz | 0.136 | 0.078 | 0.75 | 0.32 | 0.51 |
| 400-500 Hz | 0.044 | 0.034 | 0.65 | 0.56 | 0.91 |
| 500-600 Hz (swarming) | 0.027 | 0.013 | 0.75 | 0.18 | 0.34 |
| 600-1000 Hz | 0.061 | 0.036 | 0.70 | 0.44 | 0.63 |
| 1-4 kHz | 0.039 | 0.029 | 0.55 | 0.68 | 1.00 |
Values are shares of session power. Session AUC is the fraction of the twenty queenright/queenless session pairs in which the queenright session has the larger share. The last column drops CF001, a queenless session that contributes six clips yet carries the same weight as a 1 200-clip session in every session-level mean on this page. The 20-100 Hz gap, the largest in the table, is mostly that one session. The swarming band, the best of the eight, sits at p = 0.18 with all sessions and 0.34 without it. That is the same finding as above, now with a number on it.
A small convolutional network on log-mel spectrograms, with a filterbank that starts at 20 Hz so the low band is not discarded by a default. It is trained the same way three times:
- random - five stratified folds over clips. Clips from the same ten-minute file land on both sides. This is how most published results are produced.
- session - leave one recording session out.
- hive - leave one hive out. This is the only protocol that answers the question a beekeeper asks.
Folds under the last two often contain a single class, so out-of-fold scores are pooled and scored once.
| protocol | pooled AUC | balanced accuracy |
|---|---|---|
| random over clips | 0.9999 | 0.995 |
| leave one session out | 0.054 | 0.14 |
| leave one hive out | 0.375 | 0.39 |
The random number is real and means nothing. It measures whether the network can tell which recording a clip came from. The session number is the instructive one. An AUC of 0.05 is not a model that knows nothing. It is a model that confidently assigns each new session the label of the session it most resembles, and in this dataset the nearest session usually has the other label (Hive1's two days, Hive3's three). A model that had learned queens would sit near 0.5 on a held-out session at worst. This one learned sessions, and it says so.
The pooled AUC hides which sessions are wrong. From the out-of-fold
probabilities of the leave-one-session-out run (results/cv/,
src/session_checks.py, results/session_checks.json; added 2 October):
| held-out session | truth | clips | called queenright | median p(queenright) |
|---|---|---|---|---|
| CF001 | queenless | 6 | 83 % | 0.99 |
| CF003 | queenright | 600 | 13 % | 0.01 |
| CJ001 | queenless | 374 | 81 % | 0.96 |
| GH001 | queenright | 600 | 21 % | 0.07 |
| Hive1, 31 May | queenless | 600 | 50 % | 0.50 |
| Hive1, 12 June | queenright | 600 | 0 % | 3e-07 |
| Hive3, 12 July | queenless | 600 | 99.8 % | 1.00 |
| Hive3, 15 July | queenless | 600 | 95 % | 1.00 |
| Hive3, 20 July | queenright | 321 | 0.6 % | 1e-05 |
No session is called right by a majority of its clips. Seven of the nine are called wrong on more than 80 % of their clips, GH001 on 79 %, and the ninth, Hive1 on 31 May, is a coin: 301 of 600. Ranked by mean probability, every queenless session sits above every queenright one, a session-level AUC of 0 out of 20 pairs. Only 2 of the 126 relabellings of nine sessions are that extreme in either direction, p = 0.016. At the unit this page insists on, the reversal is not noise.
Two things in that table change what was said above.
The explanation covers five sessions, not nine. "The nearest session has the other label" can only be about Hive1 and Hive3, where the same hive is in the training set on another day. The four citizen-science hives have one session each, nothing of theirs is in training, and they are called wrong just as firmly (5 of 6 clips, 87 %, 81 %, 79 %). It does not fit Hive3 either: the queenless day of 12 July has a queenless neighbour three days later and a queenright one eight days later, and is called queenright on 99.8 % of its clips. The stored probabilities do not say which training session a held-out one was matched to. The sentence was a guess and should have been written as one.
Chance is not 0.5 under this protocol. Holding a queenright session out leaves a training set that is 41-45 % queenright; holding a queenless one out leaves 49-57 %. A model that learned nothing and returned its training prior would score every queenright clip below every queenless one: pooled AUC 0.00, not 0.50. The network's outputs are nowhere near those priors (medians from 3e-07 to 1.00), so the prior alone is not what it is doing. But some pull below 0.5 is built into leave-one-out with pooled scoring, and its size cannot be read from these files. A run with the labels shuffled across sessions would measure it, and that has not been done. "Near 0.5 at worst" was too strong.
Leave-one-hive-out reads the same way. The network gives a held-out hive one label whatever the day. Hive1 is called queenless on 99.8 % of its queenless clips and on 100 % of its queenright ones (within-hive AUC 0.49). Hive3 is called queenright on 75 % and 82 % (0.38). CF003 and GH001 are called wrong on 76 % and 79 %, CF001 on 6 of 6, and CJ001 is the one hive called right, on 65 %. The pooled 0.375 comes from which label each hive is handed, not from any ordering within a hive.
A hive sounds the way it does because tens of thousands of bees beat their wings. Each bee is a periodic source: a stroke at a fundamental f0 that moves with the bee's size, load and temperature. A periodic source has harmonics. The colony spectrum is written as that superposition and nothing else:
S(f) = g * [ sum_k p(f0_k) sum_h a(h, f0_k) L(f; h*f0_k, h*gamma) + B(f) ]
Here p(f0) is the distribution of wing-beat frequencies in the colony, a 55-point histogram from 80 to 350 Hz, one per clip. The harmonic profile of a stroke, a(h, f0), is a small network of harmonic number and fundamental, shared by every clip. L is a Lorentzian line whose width grows with harmonic number. B(f) is a power-law background per clip for everything that is not a wing beat. The model is fitted to the spectra alone, by log-spectral error. The queen label never enters the fit. Under the session and hive protocols the shared physics (line width, harmonic profile) is refitted on the training sessions only and frozen before the held-out clips are decomposed.
What the fit finds. A single-bee spectrum with a line width of 2.4 Hz at the fundamental, and a harmonic profile that decays over five harmonics. The learned colony distributions put most of the mass at 100-160 Hz, with the textbook 180-260 Hz band nearly empty in every session. Where the mass sits differs by session far more than by class. The queenless sessions include one at 104 Hz and, on six clips, one at 245 Hz. (Section 5.1 checks this paragraph session by session; the first sentence about the mass holds for five sessions of nine.)
What the quantities are worth. Five values per clip go into a logistic regression under the same three protocols: mean and spread of f0, share of power in the comb, background slope and level. They are compared with the eight band energies and with the network:
| reading of the sound | random | session | hive |
|---|---|---|---|
| log-mel CNN | 0.9999 | 0.054 | 0.375 |
| 8 band energies + logistic regression | 0.915 | 0.522 | 0.566 |
| wing-beat model, full p(f0) (55 numbers) | 0.909 | 0.203 | 0.387 |
| wing-beat model, 5 named quantities | 0.842 | 0.595 | 0.618 |
The pattern is the same in every row. The more freedom a reading has, the better it does on the random split and the worse it does on a new colony. The network is best on clips it has effectively seen and worst on colonies it has not. The full wing-beat histogram, 55 numbers per clip, is enough to identify sessions again and falls to 0.20. Only the five quantities with physical names stay above chance on a new session (0.60) and a new hive (0.62). That is modest, it rests on nine sessions, and it is not enough to act on.
The caveat that matters. The model knows what a harmonic series looks like. It does not know what a bee is. Any periodic source in the recording is decomposed into "bees" at whatever fundamental fits: mains hum at 50 or 60 Hz and its harmonics, a fan, a compressor in the citizen-science recordings. The mass at 100 and 150 Hz in several sessions (two, on a closer look: section 5.1) is exactly where the harmonics of a 50 Hz mains would sit. This is the same lesson as section 4 in a different coat. The decomposition is only as good as the recording. A physical model at least makes that failure visible - a distribution parked at mains harmonics can be seen in the left panel - where a spectrogram network hides it.
Two sentences above describe nine distributions at once. Opened up, as
shares of p(f0) on the 55-point grid (results/session_checks.json; added
2 October):
| session | source | 80-95 Hz | 100-160 Hz | 180-260 Hz | p at 100 Hz | comb share |
|---|---|---|---|---|---|---|
| CF001 | OSBH | 0.21 | 0.10 | 0.03 | 0.00 | 0.003 |
| CF003 | OSBH | 0.30 | 0.18 | 0.21 | 0.05 | 0.74 |
| CJ001 | OSBH | 0.84 | 0.07 | 0.03 | 0.01 | 0.51 |
| GH001 | OSBH | 0.46 | 0.42 | 0.02 | 0.03 | 0.92 |
| Hive1, 31 May | NU-Hive | 0.16 | 0.69 | 0.13 | 0.04 | 0.67 |
| Hive1, 12 June | NU-Hive | 0.03 | 0.89 | 0.03 | 0.03 | 0.90 |
| Hive3, 12 July | NU-Hive | 0.01 | 0.83 | 0.02 | 0.00 | 0.97 |
| Hive3, 15 July | NU-Hive | 0.06 | 0.60 | 0.30 | 0.15 | 0.61 |
| Hive3, 20 July | NU-Hive | 0.12 | 0.55 | 0.26 | 0.21 | 0.47 |
Correction: "most of the mass at 100-160 Hz" is true of the five NU-Hive sessions (55-89 %) and of none of the four citizen-science hives (7-42 %). In those four the distribution is piled against the bottom of the grid. The four lowest grid points, 80-95 Hz, hold 21-84 % of the mass, and the single lowest point, 80 Hz, holds 30 % in CJ001 and 31 % in GH001. A distribution leaning on the edge of its grid is asking for a fundamental the grid does not offer. CJ001's mean of 104 Hz and GH001's 127 Hz mark where the grid starts, not a wing-beat frequency. The 180-260 Hz band is nearly empty (2-3 %) in five sessions, not in every one: it holds 13 %, 21 %, 26 % and 30 % in the other four.
The mains suspicion narrows to two sessions. A line at 100 Hz that stands clear of its neighbours appears in Hive3 on 15 and 20 July: 15 % and 21 % of the mass in one grid point, 6.1 and 5.2 times the mean of the points either side. In the other seven sessions the 100 Hz point holds 0-5 % and 0.2-1.6 times its neighbours. Hive3 on 12 July, the same hive three days earlier, has 0.4 % there. So "several sessions" was two, and whatever puts a line at 100 Hz was not there, or not audible, on 12 July. Those are also the two sessions with the most mass in 180-260 Hz (30 % and 26 %).
CF001 has no comb to describe. On its six clips the model assigns 0.3 % of the power to wing beats and the rest to the background. The 245 Hz quoted above is the mean of a distribution fitted to almost nothing.
The named quantities at the session level. The 0.60 and 0.62 in the table are pooled over clips. Four of the five quantities are stored as session means (the background level is not), and they go through the same exact test as the band table in section 3:
| quantity | queenright | queenless | session AUC | exact p | p without CF001 |
|---|---|---|---|---|---|
| mean f0 (Hz) | 146 | 159 | 0.45 | 0.74 | 0.66 |
| spread of f0 (Hz) | 57 | 54 | 0.65 | 0.83 | 0.54 |
| comb share | 0.76 | 0.55 | 0.65 | 0.36 | 0.63 |
| background slope | 1.40 | 2.03 | 0.35 | 0.36 | 0.69 |
None separates nine sessions (p from 0.36 up), and the two that come closest lose even that without the six-clip session. Whatever the regression finds across clips, it is not a difference between session means.
Which clips. The network rows use 4 301 clips (at most 600 per
session). The three logistic-regression rows use 2 406 (at most 300 per
session, MAX_CLIPS_PER_SESSION in src/wingbeat_pinn.py). Each row is
comparable across its three protocols. The rows are not on the same clips.
Eight checks, all passing:
- the queen label, hive and session are parsed from every file-name pattern in the dataset, and documentation files are rejected
- only segments annotated as hive sound are kept; a comma decimal is read correctly
- the mel filterbank covers every frequency from 40 Hz up, so the band where hive sound lives is not silently discarded
- a single bee in the wing-beat model produces peaks at f0, 2f0, 3f0, 4f0 and nowhere else
- the line width at the third harmonic is three times the width at the first
- a colony placed entirely at 250 Hz is read back as f0 = 250 Hz with zero spread
- the result is pinned: the random-split AUC must exceed 0.99, the network's session and hive AUCs must be below 0.5, and the five wing-beat quantities must land between 0.5 and 0.8 on both leave-out protocols. The page is not allowed to claim a working detector.
- no band separates the classes: the exact permutation test over the 126 relabellings of nine sessions returns p > 0.1 for every band, with and without the six-clip session. The estimator itself returns 1/126 on a perfectly separated case.
A second file, tests/test_readme_numbers.py (twelve checks, added
2 October), recomputes sections 4.1 and 5.1 from the stored outputs and
pins every number in them to results/session_checks.json. It needs no
audio.
- Nine sessions from six hives. Every cross-colony number above has a sample size of six, not four thousand. The 0.60 / 0.62 could be 0.5 on the next six hives.
- Two recording projects, two microphones, two countries. Part of what separates sessions is equipment. A dataset with one microphone in twenty hives would answer the hive question far better than this one can.
- Queen removal is not the only thing that changed between the NU-Hive sessions. Date, weather, time of day and forage did too. The label is the queen; the sound is everything.
- The wing-beat model is a spectral model, not a bee counter. It cannot tell a bee from any other periodic source, and it does not model the box, which shapes the spectrum below 300 Hz.
- No swarming, no disease, no robbing. The dataset labels one state.
This started as a dormant folder called beehive-acoustics. The code was
written on 7 September, the download stopped half way, and nothing was ever
analysed. Fetching the rest of the raw audio (3.7 GB) from Zenodo only
worked with resumable curl -C - and retries. The finished version, with
the wing-beat model added, went public on 12 September. The exact
permutation test in section 3 and the eighth check came later the same day.
Until then the band claim rested on eyeballing the spread. On 2 October
the pooled numbers of sections 4 and 5 were opened session by session
(4.1, 5.1). That cost the page one explanation and half of one sentence,
both marked where they stand. CHANGELOG.md keeps the dated list.
Two things I would not call finished. The global fit of the wing-beat model
reaches a log-spectral RMSE of 0.60 after 1 500 Adam steps, which is
mediocre. And the mains-harmonic problem in section 5 is stated, not
handled. Notching or masking the 50 Hz harmonics before the fit, and running
more steps, are the obvious next moves. For scale, train.py takes about
25 minutes on a laptop GPU and wingbeat_pinn.py about 20 minutes on a
laptop CPU.
The physics core is public in this repository:
src/prepare.py- the recordings to labelled clips, with provenancesrc/spectra.py- session-level spectra and the band tablesrc/model.py- the log-mel front end (filterbank written out) and the networksrc/train.py- the three protocols, pooled scoring, resumablesrc/band_permutation.py- the exact permutation test on the session band table (standard library only)src/wingbeat_pinn.py- the wing-beat mixture model, its named features, and the protocol comparisonsrc/session_checks.py- sections 4.1 and 5.1: the stored outputs read one session at a time (numpy only, no audio)tests/test_all.py- the eight checks abovetests/test_readme_numbers.py- sections 4.1 and 5.1 pinned to the results files
results/ holds every number on this page as JSON. data/SOURCE.md says
how to fetch the recordings; they are not redistributed here.
pip install -r requirements.txt
python src/download.py && python src/prepare.py
python src/spectra.py && python src/band_permutation.py
python src/train.py # ~25 min on a laptop GPU
python src/wingbeat_pinn.py # ~20 min on a laptop CPU
python src/session_checks.py # seconds, no audio
python src/figures.py && python -m pytest tests
Documentation, figures and result files: CC BY 4.0. Source code in src/
and tests/: MIT.
The recordings belong to their authors and are distributed through Zenodo under CC BY 4.0. Cite them, not this page:
- Inês Nolasco, Emmanouil Benetos. 2018. To bee or not to bee: An annotated dataset for beehive sound recognition. Zenodo, doi:10.5281/zenodo.1321278. Recordings from the NU-Hive project and the Open Source Beehives project.
Other acoustic projects built on the same physics-first approach are listed on the profile README.