A telco fund paid for by old network gear

One New Zealand announced the Nature Tech Fund on 17 November 2026. The money comes from an unusual place. When One NZ retires network equipment and other hardware, its recycling partner Echo Tech recovers materials from it, and the rebates from that recycling go into the fund. One NZ says it processed more than 90 tonnes of e-waste in its latest financial year and kept 99% of it out of landfill. The fund extends the company's National Conservation Partnership with the Department of Conservation (DOC).

We couldn't find a dollar figure for the fund, an application process or a timeline in the launch coverage. What we do know is that three projects got backing at launch:

  • Hark's Fiordland Piopio Expedition. Acoustic monitors in remote Fiordland forest will record continuously and send data over One New Zealand Satellite. AI models will then check the audio for any trace of the piopio, a bird currently listed as extinct.
  • A connected camera trial with the Predator Free Hauraki Coromandel Community Trust, using 4G and 5G links.
  • Urban biodiversity monitoring at One NZ's Auckland support office, where cameras, sensors and AI will watch over 5G-connected beehives.

Nicky Preston, One NZ's head of sustainability and corporate affairs, said the company was drawn to "the circular thinking behind it." She also said that "Technology can't replace the local people who know their whenua and species best," and described the tools as extra eyes and ears for those people in places that are hard to reach.

The piopio project is the one most worth a closer look for AI builders, because it puts an open foundation model on a small field device.

What HARK is and how it works

HARK comes from the 800 Trust, an ecological charity based in Taranaki. The Toi Foundation gave the trust its first development grant, which paid for materials and a proof-of-concept unit. The trust now builds the devices in Stratford. Its project lead is Miranda Wells.

Predator Free NZ profiled the device in August 2026. Here's the hardware it described:

  • three microphones that listen all the time at sample rates up to 192 kHz
  • sensors for temperature, humidity and barometric pressure, so changes in animal activity can be matched against storms and other weather
  • a solar panel tuned for low light under the forest canopy
  • a stainless steel body built to survive kea and rats, with a design life of ten years
  • a 4G uplink, falling back to a mesh network between units where there's no cell coverage
  • automatic setup, so volunteers don't need IT skills to deploy it

Most acoustic recorders used in New Zealand conservation write raw audio to an SD card. Someone has to collect the card and process the files weeks later. HARK classifies sound on the device and sends out only the results. Magica reports that a detection triggers a phone alert, so a team member can play the clip and decide whether to act, even in the middle of the night. Several units can work together to estimate where a sound came from by triangulation. Magica puts the unit at roughly the size of a one-litre milk bottle.

According to Wells, as quoted by Predator Free NZ, the device "can run over 100 AI models in real-time." The model it names is Google's Perch v2. The 800 Trust was among 16 Asia-Pacific groups picked for Google DeepMind's AI for the Planet Accelerator, and HARK won the public-good category at the 2026 NZ Hi-Tech Awards.

The model underneath: Perch 2.0

Google DeepMind released Perch 2.0 as an open model on Kaggle in August 2025. The first Perch came out in 2023 and was downloaded more than 250,000 times, according to DeepMind. The new version was trained on nearly twice as much data. It also goes beyond birds to cover mammals, amphibians and human-made noise.

The Perch 2.0 paper, by Bart van Merriënboer, Tom Denton and colleagues, gives the technical detail:

  • Input: mono audio at 32 kHz, cut into five-second windows and turned into log mel-spectrograms with 128 bins between 60 Hz and 16 kHz.
  • Backbone: EfficientNet-B3, about 12 million parameters. Perch 1.0 used the smaller B1.
  • Output: a 1,536-dimension embedding for each window, plus class scores.
  • Training data: about 1.54 million labelled recordings from Xeno-Canto, iNaturalist, the Tierstimmenarchiv and FSD50K. Birds make up roughly 1.37 million of them. There are 14,795 classes in total.
  • Results: on BirdSet it reaches 0.908 AUROC, up from 0.839 for Perch 1.0, and it also beats its predecessor on BEANS. Neither result involved fine-tuning the embedding model.

HARK records at up to 192 kHz, far above Perch's 32 kHz input, so the audio has to be resampled before Perch can classify it. That higher rate is probably there for bats and other ultrasonic sounds. The trust hasn't published how it splits the work between models.

In Australia, DeepMind credits the original Perch with helping BirdLife Australia and the Australian Acoustic Observatory build classifiers that led to a newly found Plains Wanderer population. That's a useful precedent for a search like the piopio expedition.

The limits we can see

A rediscovery search is about as hard as acoustic AI gets. Most of the audio contains nothing of interest, the target may not exist, and a single positive would be big news. That means false positives carry a real cost.

AUT published relevant New Zealand research on 24 August 2026. Dr Akbar Ghobakhlou's team found that small possum detectors trained in the usual way raised hundreds of false alarms on native forest recordings that had no possums in them. Their fix was to run BirdNET over possum calls, see which bird species it confused them with, and feed those birds' calls back in as "hard negatives." False alarms fell sharply while detection stayed strong. The gain showed up only on recordings the models hadn't seen during training, and AUT didn't publish exact error rates.

There are some other gaps:

  • Reference audio. A classifier needs examples of the target sound. Nothing we found says whether Perch's training sources hold usable piopio recordings, so any detection will need careful human review.
  • No independent accuracy figures. The 800 Trust hasn't released detection rates, false-alert rates or a full model list.
  • Deployment scale. One NZ hasn't said how many devices will go to Fiordland or for how long.
  • Price. Magica reports about $1,000 per unit, which the trust itself says is more than many community groups can afford.

The first independent test is under way. Taranaki Regional Council is comparing HARK with AI cameras from Critter Solutions and 2040 through 2026 in the Zero Possum area between Oākura and Ōkato. Magica reports this is part of three-year research that DOC is funding with $1.5 million. A scientist designs and evaluates the study, and the council may even release a possum into a cleared block to see which devices pick it up.

What builders can do now

You don't need a HARK unit to try this approach. Google's perch-hoplite library is released under Apache 2.0 and is built for an embed, search and classify workflow, which DeepMind calls agile modelling. You embed your audio once, search it with a single example call, label what comes back, and train a light classifier on top. The README says new bioacoustic classifiers often take under an hour.

Setup is short. Install libsndfile1 and ffmpeg with apt, then pip install perch-hoplite. Perch and BirdNET need TensorFlow, so add the [tf] extra for CPU or [tf-cuda] for a GPU. Start with the notebooks agile/01_embed_audio.ipynb and agile/02_agile_modeling.ipynb. The older google-research/perch repo warns that parts of it are out of date and points inference users to hoplite. It still holds export code for TensorFlow Lite, which is a useful starting point if you're targeting low-power hardware.

Councils, trusts and landowners around the country are sitting on hard drives full of audio that nobody has analysed. We think a sensible first project is to embed one season of it, search for a familiar call such as ruru or possum, and then apply AUT's hard-negative method before putting anything on a device. If you build something along these lines, tell us about it through the ASAI projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member