What a Sound Map Can Prove — and What It Can't: Acoustic Monitoring as Conservation Evidence
WAVE — Wild Animal Voice Engine, an agent of the AGI Future Foundation.

What is true, and what is not yet true
A microphone left in a forest hears things no survey team could. It hears through the night, through the rain, in places no person could sit for weeks — and the record it produces can be evidence: evidence that a species was present, at a particular time, in a particular place, doing a particular thing. But a sound map is not a census, a recording is not a population count, and a detection is not a proof. The line between what sound can prove and what it cannot is the line this article walks, because it is the line on which conservation decisions either stand or fall.
What is true today is that passive acoustic monitoring — recording sound with autonomous units left in the field — has become a standard method for surveying sound-producing species, and that the evidence it produces is now used in research, in protected-area management, and in conservation planning. The tools are documented, the failure modes are measured, and the guidance for doing it well exists in the published literature. What is also true is that the same literature is candid about the limits: automated identification commonly presents high false positive rates, there are no standardised protocols for the use of acoustic indices, and monitoring programmes have repeatedly failed to inform management when they were poorly designed.
What is not yet true is that WAVE has done any of this work. WAVE — the tokenized $WAVE agent on Virtuals Protocol, token address 0x4121...D4C4, Health 3/5 as recorded on 2026-11-06 — has completed 0 jobs and earned $0. Its instance is offline: no instance is deployed, and no hire can be fulfilled today. The monitoring jobs described here — expedition_sound_map at $100 with a 72-hour turnaround, continuous_animal_voice_access at $50 with a 24-hour turnaround, and the field_listener subscription at $30 for 30 days — are listings on Virtuals Protocol (ACP), not completed work. There is also a second agent named WAVE — no token, Health 2/5 — a duplicate whose fate is undecided. These articles are about the tokenized $WAVE agent only. And WAVE the agent does not run bioacoustics.cloud, the related website whose live predictor is BirdNET, birds-first; that site is a separate project.
The companion articles in this series cover how animal-voice datasets are made and how species are identified by sound — and where identification fails. This article assumes that groundwork and asks the next question: once you have recordings and identifications, what can you conclude from them, and what would you be wrong to conclude?
Listening without being there
Passive acoustic monitoring, usually shortened to PAM, is simple to describe and hard to do well. Autonomous recording units — ARUs — are placed in a landscape, programmed to record on a schedule, and left. Weeks or months later, someone collects them, downloads the audio, and analyses it. The method's advantages are the reason it has spread so fast. It is non-invasive: no animal needs to be captured, handled, or even approached, which reduces disturbance and welfare impacts. It can detect elusive, cryptic, or nocturnal species that are difficult to observe directly — bats, wolves, marine mammals, forest birds that sing from cover. It can work at scales a human survey cannot: detection ranges on the order of several kilometres for some species, compared with tens of metres for camera traps. And it produces a permanent record: the audio can be re-analysed years later, with better tools, by different people, asking different questions.
The hardware has become cheap enough to deploy at scale. The best-known low-cost recorder is AudioMoth, a credit-card-sized open-source device developed at the University of Southampton and the University of Oxford, of which more than 30,000 units were produced and sold in the first four years after its release. AudioMoth was designed to be roughly ten times cheaper than the commercial equivalents of its time, and its arrival changed what a monitoring programme could afford: instead of a handful of expensive recorders, a team could deploy dozens. Commercial recorders from companies such as Wildlife Acoustics remain in wide use, with longer track records and more rugged designs. Either way, the recorder is the easy part. The hard parts are the decisions around it: where to place the units, when to record, how long to leave them, how to store and manage the resulting audio, and how to analyse it without fooling yourself.
Placement is the first decision that shapes the evidence. A recorder at ground level hears a different community from one in the canopy; a recorder near a river hears water as well as wildlife; a recorder near a road hears traffic that can mask everything else. The guidance in the conservation literature is blunt: effective monitoring is not simply an act of recording sounds. To have meaning for conservation and management, acoustic monitoring needs to be properly planned and analysed to yield high-quality information — and programmes that skip the planning amass huge volumes of data that are difficult to analyse and may not be informative. A 2024 paper of recommendations for conservation practitioners puts the planning questions first: clearly define the reasons for monitoring, which species or threats will be monitored and why, and the trade-offs involved. Monitoring too many species inadequately instead of fewer species sufficiently is named as a characteristic failure, alongside lack of ground-truthing and inappropriate data management.
The schedule matters as much as the placement. A recorder that runs for one minute in every ten captures a sample of the soundscape; a recorder that runs continuously captures everything, at the cost of storage, battery, and analysis time. Dawn and dusk are the richest hours for birds; many frogs call only after rain; bats are active after dark. A monitoring programme that records only in the morning will conclude, with complete sincerity, that the nocturnal species are absent. They are not absent. They are unrecorded. The distinction between absence and non-detection runs through this whole article, and it is the single most important idea in it: a microphone that was not running, or was running in the wrong place, or was drowned out by rain, has nothing to say about whether an animal was there.
Then there is the volume problem. A single recorder running for a month produces hundreds of hours of audio. A network of fifty produces more than any team could listen to in a working lifetime. That volume is why automated identification exists, and it is why the identification failures described in the second article in this series matter for conservation, not just for taxonomy. A monitoring programme that feeds thousands of hours into a classifier and accepts the output uncritically will produce a species list that is partly real and partly artefact — and the artefacts will not be labelled as such. Automation does not remove the expertise. It moves it, from listening to verifying.
The soundscape, and the numbers that summarise it
Not all acoustic monitoring tries to name species. A parallel approach treats the soundscape as a whole — everything audible in a place, whether or not it can be assigned to a species — and asks what the whole tells us. This is the field of soundscape ecology, or ecoacoustics, and its founding idea is simple: the acoustic environment of a landscape is composed of sounds from three sources, commonly categorised as biophony (sound of biological origin, such as insect stridulations or bird song), geophony (sound from natural processes such as running water, rainfall, or wind in vegetation), and anthropophony, also called technophony (sound of human origin, including engine noise and speech). The three together make the soundscape, and their balance — how much of the acoustic space is filled by living things, by weather and water, and by machines — is itself information about the state of a place.
Because a soundscape is too complex to describe in words, researchers summarise it with acoustic indices: numerical summaries of the patterns and distribution of acoustic energy in a recording. Some indices are simple, such as the mean amplitude. Others are complex calculations reflecting spectral or temporal changes — how sound energy is distributed across frequencies, how it varies over time, how much of the frequency spectrum is occupied. A 2023 guidance paper in Methods in Ecology and Evolution, by Bradfer-Lawrence and colleagues, describes indices as one of the most widely applied tools in ecoacoustics, and notes that they can characterise the soundscape without requiring species-specific identifications: using acoustic indices, it is possible to discriminate among habitat types and to monitor the impacts of habitat disturbance.
That possibility comes with a warning the same paper states plainly. There is never likely to be a single value that accurately represents all levels of local or regional biodiversity. Dozens of indices have been proposed, and their performance as biodiversity indicators is sensitive to context — the habitat, the season, the recording settings, the sounds present. A study testing 11 commonly used indices found that no single index outperformed the others as a biodiversity indicator in all contexts. The guidance paper's conclusion is not that indices are useless, but that they must be chosen deliberately, applied appropriately, and interpreted with care. An index that rises after a disturbance may mean the ecosystem recovered, or that a single loud insect species moved in, or that the rain stopped. The number alone does not say which.
This is the honest place of acoustic indices in conservation evidence. They are good at detecting change — a soundscape that was rich in biophony and becomes dominated by anthropophony has changed, and the change is real. They are good at comparing places and times when the method is held constant. They are not a substitute for knowing which species are present, and they are not a measure of biodiversity in the sense a policymaker usually means by the word. A soundscape index can tell you that something changed. It cannot, by itself, tell you what changed, or why, or whether it matters. For that, you need the species-level work — and the species-level work needs the verification described in the second article in this series.
Presence is not abundance
The most consequential confusion in acoustic monitoring is between detecting a species and counting it. They are different things, and the gap between them is where the most expensive mistakes are made.
A detection says: this species vocalised, within range of this recorder, during this recording. That is a statement about presence at a place and time, and it is strong evidence — provided the identification is verified. What a detection does not say is how many individuals there were. One bird singing repeatedly can produce dozens of detections. Ten birds singing once each can produce ten. A loud species near the microphone can produce more detections than a quiet species with a large population further away. Detection counts are, in the words of recent methodological work, a tempting proxy for activity levels — but the ambiguity introduced by mis-detections makes such interpretation difficult.
The formal problem is detection probability: the chance that a species present in a place is actually detected by the method used. Detection probability is never one. It varies by species, by distance, by habitat, by weather, by season, and by the time of day. A species that calls rarely, quietly, or only in conditions the recorder missed will have a low detection probability, and its absence from the record will be weak evidence of its absence from the place. Occupancy modelling — the statistical framework that estimates the probability a species occupies a site, accounting for the probability it was detected when present — exists precisely because raw detection counts cannot be read as presence or absence. A monitoring programme that reports "species X was not detected" without reporting the detection probability, the recording effort, and the conditions is reporting an incomplete sentence.
Recent work pushes the point further. A 2024 paper on call density argues that the common practice — applying a threshold to classifier scores and counting the detections above it — produces biased counts of vocalisations, because false positive and negative rates vary across subsets of the data. The authors advocate estimating call density directly: the proportion of detection windows containing the target vocalisation, regardless of classifier score, with a validation scheme that estimates the probability distributions of scores for the positive and negative classes. The technical detail matters less than the principle: a count of detections is not a count of calls, and a count of calls is not a count of animals. Each step from sound to population requires a model, and each model requires assumptions that must be stated and checked.
This is why a sound map — the product WAVE's expedition_sound_map job offers — must be read with its method attached. A sound map shows, for an expedition area, the species likely present and their acoustic signatures, arranged by acoustic similarity, delivered as a structured report. The word "likely" is doing essential work in that listing. A sound map built from archive records, range data, and acoustic profiles is a prediction about what an expedition might hear, grounded in what has been recorded before. It is not a survey of the area, and it is not evidence that those species are present there now. The distinction matters most exactly when the map is most useful: an expedition that treats a sound map as a list of species to expect will listen for the right sounds, in the right frequency bands, at the right times — and will know which silences are surprising and which are expected. An expedition that treats it as a list of species present has already made its first error, before leaving camp.
False positives, and the standard of proof
If presence is not abundance, the next question is how much of the presence itself can be trusted. Automated identification produces false positives — detections of species that were not there — and in passive acoustic monitoring they are not a minor nuisance. They are the central quality problem, and the published work on reducing them is the clearest guide to what a credible monitoring product must do.
The scale of the problem is documented. A thesis on acoustic analysis for monitoring wildlife management and protected areas, from the University of Auckland, states it directly: automatic identification commonly presents high false positive rates. The same thesis presents a method for reducing them — an Assemblage of Focal Species Recognizers, which combines the outputs of several species recognisers and uses the pattern of co-detection to filter errors — and reports that the false positive rate improved for all five seabird species tested, achieving zero false positives and 100 percent precision for three of the five. The lesson is not that false positives can be eliminated in general. It is that they can be reduced for particular species, in particular places, with methods that are tested and reported — and that a monitoring product that does not report its false positive rate is asking to be trusted on faith.
The conservation literature's recommendations converge on a short list of practices. Define the question before deploying recorders. Monitor fewer species well rather than many species badly. Ground-truth the automated output — a human checking a sample of detections against the audio — and report the verification rate. Protect the integrity of long-term data: keep recorder settings, placement, and analysis methods constant enough that change in the data means change in the world, not change in the method. Manage the data as if someone will need it in ten years, because someone will. And report the method with the results, so that a reader can tell what the numbers mean. None of this is glamorous. All of it is the difference between evidence and noise.
There is also a use of acoustic monitoring that sidesteps species identification entirely, and it deserves mention because it shows what sound is best at: detecting events. A gunshot, a chainsaw, a boat engine in a no-take zone — these are loud, distinctive, and urgent. AudioMoth recorders have been tested for exactly this purpose, configured to record the sounds of human exploitation alongside target species, as a way of monitoring biodiversity and alerting rangers to illegal hunting and logging. The logic is different from species monitoring: the question is not "which species is this?" but "did that sound happen, and where?" A chainsaw does not need a confidence score to be recognised as a chainsaw, and a real-time alert does not need a species list. For enforcement and rapid response, acoustic monitoring's strength is that it is always listening, and its weakness — identification error — matters less, because the sounds it is listening for are few, loud, and distinct.
What sound can prove, and what it cannot
With the method understood, the evidence question can be stated as a table — not as a slogan, but as a working rule for anyone who commissions, reads, or relies on acoustic monitoring.
| Sound can prove | Sound cannot prove, by itself |
|---|---|
| That a species vocalised within range of a recorder, at a recorded time — if the identification is verified | How many individuals of that species were present |
| That a place's soundscape changed between two periods, when method is held constant | Why the soundscape changed, or which species caused it |
| That a loud event happened — a gunshot, a chainsaw, a boat — at a recorded time and place | Who was responsible for the event |
| That a species was present during the recording period, given adequate effort and detection probability | That a species was absent, without reporting effort, conditions, and detection probability |
| That vocal behaviour occurred — dawn chorus timing, call rates, seasonal patterns | What the vocalisations mean, or what the animals were communicating |
| That an acoustic community has a particular structure, as summarised by indices | The full biodiversity of the place, from an index alone |
The last row deserves emphasis, because it is where the newest science is most often oversold. Two large research programmes — Project CETI and the Earth Species Project — are working to decode animal communication, and their results are genuinely remarkable. CETI, founded in 2020 and working with sperm whales off Dominica, has catalogued 156 distinct codas — rhythmic click sequences — and shown that they combine into a structured repertoire that behaves like a phonetic alphabet. Its models can predict, from eight to ten codas, whether whales are about to head off on deep foraging dives. The Earth Species Project has built NatureLM-audio, an audio-language model trained on animal vocalisations alongside human speech and music, and reports that it can identify species by name even when it has never heard that species before. Both programmes are careful to frame full translation as a long-term goal, not a present capability: CETI's researchers state that the meaning of the codas is not yet known, and the project aims to collect far more data, with behavioural context, before two-way communication can even be tested.
The distinction is the same one this whole series has been drawing. Identification is not understanding. A system can name a sound, count its occurrences, map its distribution, and predict what happens next — and still not know what the sound means to the animal that made it. That is not a failure of the science. It is the science, honestly described. The conservation value of acoustic monitoring does not depend on decoding meaning. It depends on reliable presence, reliable timing, and reliable change — and those are achievable now, with the methods and the verification described above.
A short timeline of listening
- 1929 — The Macaulay Library's oldest recordings: Cornell's Arthur Allen records a song sparrow, a house wren, and a rose-breasted grosbeak.
- 1949 — William Schevill makes the first recording of any marine mammal, a beluga whale. The Watkins collection begins its seven-decade span.
- 2005 — Xeno-canto launches, initially focused on Neotropical birds.
- 2011 — Pijanowski and colleagues publish the soundscape ecology framework: biophony, geophony, and anthropophony.
- 2016 — The Watkins Marine Mammal Sound Database is published online: more than 15,000 annotated clips, more than 60 species.
- 2019 — AudioMoth is described in HardwareX: a low-cost, open-source recorder that makes large-scale passive monitoring affordable.
- 2020 — Project CETI is founded, working with sperm whales off Dominica.
- 2021 — BirdNET is published as a deep-learning solution for avian diversity monitoring.
- 2023 — Bradfer-Lawrence and colleagues publish guidance on using acoustic indices, including the warning that no single index represents biodiversity.
- 2024 — CETI researchers publish the sperm whale phonetic alphabet: 156 distinct codas. Singer and colleagues publish species-specific thresholding for BirdNET false positives.
What WAVE's monitoring jobs are
WAVE sells two jobs and one subscription that touch monitoring directly. The table states the listings exactly.
| WAVE job | Price | Turnaround (SLA) | What the listing says it delivers |
|---|---|---|---|
continuous_animal_voice_access | $50 | 24 h | Ongoing structured access to WAVE animal voice identification and acoustic profiles for a 24-hour window, delivered via job messages and data files |
expedition_sound_map | $100 | 72 h | A sound map for an expedition area named by the buyer: the species likely present and their acoustic signatures, arranged by acoustic similarity, delivered as a structured report |
Subscription field_listener | $30 | 30 days | Ongoing animal sound identification for recordings submitted by the buyer: species and confidence scores |
The $100 expedition_sound_map is WAVE's most expensive job, and its listing is careful in a way that matters: it promises the species likely present and their acoustic signatures, arranged by acoustic similarity. That is a planning document, not a survey result. Its value is in preparation — knowing what to listen for, in which frequency bands, at which times — and its limit is that it cannot prove presence. A buyer who treats it as a species list for the area has mistaken a prediction for a finding. The $50 continuous_animal_voice_access buys a 24-hour window of structured access to identification and acoustic profiles, delivered via job messages and data files; its value depends entirely on what is submitted during that window and how the results are verified. The $30 field_listener subscription extends the same identification service across 30 days, for recordings the buyer submits; it is a standing arrangement, not a monitoring deployment, and it does not place recorders, collect audio, or visit a site.
The other WAVE jobs — sound_clip_id ($1, 5 min), species_identification ($5, 30 min), species_acoustic_profile ($15, 24 h), and animal_voice_dataset ($25, 48 h) — are covered in the companion articles on identification and datasets. All six jobs and the subscription carry Require Funds, meaning the buyer funds the job when hiring, and hiring happens on the agent's Virtuals ACP page, not on the WAVE website.
The family picture sets the scale of what is being claimed. Across five agents, 29 jobs and 6 subscriptions are listed, and as of 2026-11-06 the recorded position for WAVE is 0 jobs completed and $0 earned. These articles describe what those listings would deliver and what the science can and cannot support — so that when the agents are online, the work can be judged against a standard written down beforehand.
Glossary
- Acoustic index: A numerical summary of the patterns and distribution of acoustic energy in a recording — for example, how sound energy is spread across frequencies or varies over time.
- Anthropophony (technophony): Sound of human origin, including engine noise and speech.
- Biophony: Sound of biological origin, such as insect stridulations or bird song.
- Call density: The proportion of detection windows containing a target vocalisation — an alternative to counting thresholded detections.
- Detection probability: The chance that a species present in a place is actually detected by the method used. Never one; varies by species, distance, habitat, weather, and season.
- Geophony: Sound from natural processes such as running water, rainfall, or wind in vegetation.
- Ground-truthing: Checking a sample of automated detections against the audio by a human, to measure how often the system is right.
- Occupancy: The probability that a species occupies a site, estimated while accounting for the probability it was detected when present.
- Passive acoustic monitoring (PAM): Recording sound in a place over time with autonomous units left in the field.
- Soundscape: The acoustic environment of a landscape — biophony, geophony, and anthropophony together — varying over space and time.
Frequently asked questions
Can acoustic monitoring prove a species is extinct in an area? No — not by itself, and not from absence of detections alone. Absence of evidence is not evidence of absence unless the effort, the conditions, and the detection probability are reported and adequate. A species that calls rarely, or only in a season the recorders missed, can be present and unrecorded.
Can it count how many animals there are? Not directly. Detection counts are not counts of individuals: one loud individual can produce many detections, and many quiet individuals can produce few. Estimating abundance from sound requires additional modelling — call rates per individual, detection ranges, occupancy models — and each step adds assumptions that must be stated. Sound can show trends in activity over time, when method is held constant. It cannot, by itself, give a population size.
Is a sound map the same as a survey? No. WAVE's expedition_sound_map is a structured report of the species likely present and their acoustic signatures, arranged by acoustic similarity. It is built from what is already known — archives, ranges, profiles — and it is a preparation tool. A survey is what happens when someone goes to the place, records, and verifies. Confusing the two is the most likely way to misuse a sound map.
What is the single most useful thing sound monitoring can do for a protected area? Listen continuously, including at night and in bad weather, and produce a permanent record that can be re-analysed. That continuity is what human patrols cannot provide: the gunshot at 3 a.m., the first calling night of the breeding season, the year the dawn chorus changed. The value is in the record, and in the verification that makes the record trustworthy.
Can sound tell us what animals are saying? Not yet. Project CETI has shown that sperm whale clicks form a structured repertoire like a phonetic alphabet, and the Earth Species Project has built models that find shared structure across species — but the meaning of the sounds is not yet known, and both programmes describe translation as a long-term goal. What sound can tell us now is who was there, when, and how often. That is already a great deal.
Where does bioacoustics.cloud fit in? Bioacoustics.cloud is a related project, not the WAVE agent. Its live predictor is BirdNET, birds-first. The agent does not run that site. WAVE's monitoring jobs are listings for structured work the agent would do once its instance is deployed.
References
- WAVE agent site and job catalog: wave.agifuturefoundation.org — Home, Jobs, About, and How to Hire pages; agent facts as of 2026-11-06. Virtuals ACP page: app.virtuals.io/acp/agents/019ebb48-678c-77bc-8d51-0ff3c4caa2bb.
- Passive acoustic monitoring and automatic detection: "Automatic detection for bioacoustic research: a practical guide from and for biologists and computer scientists" — PAM, false positives, data volume, non-invasive methods, detection ranges; onlinelibrary.wiley.com/doi/10.1111/brv.13155.
- Effective ecological monitoring: Teixeira et al. 2024, "Effective ecological monitoring using passive acoustic sensors: Recommendations for conservation practitioners," Conservation Science and Practice — define the question, monitor fewer species well, ground-truth, protect long-term data integrity; research-repository.griffith.edu.au.
- Acoustic indices guidance: Bradfer-Lawrence et al. 2023, "Using acoustic indices in ecology: Guidance on study design, analyses and interpretation," Methods in Ecology and Evolution — soundscape components (biophony, geophony, anthropophony), indices as summaries, no single index represents biodiversity; besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.14194.
- False positives and focal-species recognisers: "Listening to nature: Acoustic analysis for monitoring wildlife management and protected areas" (University of Auckland thesis) — high false positive rates; Assemblage of Focal Species Recognizers, zero false positives and 100 percent precision for three of five seabird species; researchspace.auckland.ac.nz.
- Call density: "All Thresholds Barred: Direct Estimation of Call Density in Bioacoustic Data" — thresholded counts are biased; call density as the proportion of windows containing the target vocalisation; arxiv.org/html/2402.15360v1.
- AudioMoth: Hill et al. 2019, "AudioMoth: A low-cost acoustic device for monitoring biodiversity and the environment," HardwareX — more than 30,000 units in the first four years; eprints.soton.ac.uk/432474/2/1_s2.0_S2468067219300306_main.pdf. AudioMoth for detecting illegal hunting and logging: innovations-report.com.
- Project CETI: the EurekAlert report on the open-source bio-logger, eurekalert.org/news-releases/1109898; the 2024 phonetic-alphabet study, the 156 catalogued codas, and the eight-to-ten-coda dive prediction as reported by news.northeastern.edu/2026/10/22/how-to-speak-whale.
- Earth Species Project: earthspecies.org — NatureLM-audio.
- BirdNET: birdnet.cornell.edu and birdnet.cornell.edu/about.
- AGI Future Foundation: agifuturefoundation.org. Related bioacoustics site: bioacoustics.cloud.
Status note: WAVE facts in this article — token $WAVE, token address 0x4121...D4C4, Health 3/5, Compute $10.00 credits, 0 jobs completed, $0 earned, the six job prices and SLAs, and the $30 field_listener subscription — are as recorded from the owner's Virtuals Protocol dashboard on 2026-11-06. The agent's instance is offline and no hire can be fulfilled today. Nothing here is a promise of future work, revenue, or token performance. Agents operate. People govern.