China Is Teaching Machines to Hear Birds
On a still morning at Shuanggui Lake, a national wetland park outside Chongqing, three small recorders listen to a chorus no human surveyor could follow. By mid-morning the day's arrivals have already been logged: Light-vented Bulbuls calling jiu-jiu, White-browed Laughingthrushes answering with a sharper ji-ya, a Long-tailed Shrike rasping in the background. Nobody wrote those names down. Software did.
The devices are part of Soundscape China, a research project launched this year by the Institute of Zoology at the Chinese Academy of Sciences, the China Meteorological Administration's detection centre and partner institutions. Recorders are being placed around the country to collect birds, insects and frogs into one open soundscape database. Since trial operation began in July, the three units at Shuanggui Lake have logged more than 40 bird species and over 3,000 individual recordings — proof that a wetland can be surveyed by ear, around the clock, without a person in the reeds.
A fingerprint made of sound
“Just as a human fingerprint is unique, a bird's call has its own voiceprint,” says Yu Xianhuai, a senior engineer at the Liangping wetland protection centre. Identification works on three properties of the sound wave: its rise and fall, its frequency, and its rhythm. The hard part is doing that inside a noisy soundscape, where wind, water and traffic sit on the same recording as the bird.
What makes it possible is the library. “The richer the database, the more samples it holds and the more precise the annotations, the better the algorithm learns to compare — and the more reliable 'identifying a bird by its voice' becomes,” Yu says. Libraries are built two ways: by pulling audio from open sources to bulk up the sample count, and by sending people into the woods with recorders. As the library grows, the model keeps re-listening, extracting frequency, rhythm and pitch, checking its answer against annotated recordings, and adjusting itself.
The system in use at Shuanggui Lake can now recognise more than 1,500 Chinese species and over 400 foreign ones, with an overall accuracy of about 85%.
The other advantage over traditional fieldwork is stamina. A researcher with binoculars is limited by daylight, weather and attention; the devices work through the night. When a unit hears a bird it starts recording, stores the audio on a card, and pushes it over a 4G link to Beijing, where software separates, screens and analyses the calls.
Why a wetland's sound matters
Wetland birds are indicator species: their fortunes stand in for the health of the whole system. Compare voiceprints across seasons and you can read migration timing, habitat preference and population change — a cheap early warning of a wetland in trouble. The project's organisers also see a public future in the data: soundscape tours, slow live-streams and listening checkpoints that let visitors hear what monitoring hears.
The same idea, one province over
In Jiangsu, a parallel programme has been running across forests, wetlands, lakes and hills in ten cities, including Nanjing, Suzhou and Wuxi. There, AI video recognition and AI sound recognition work as “electronic eyes and ears”. A unit in Nanjing's Lüshuiwan urban wetland filters out wind, rain and flowing water on the spot, then captures bird calls within about 100 metres of the tree it hangs in.
Its training material is local: open-source audio, years of Jiangsu survey records, field photographs and recordings of local birds, assembled into a provincial library of more than 400 sound vouchers — roughly 510,000 recordings — plus over 500 image vouchers and 100,000 photographs. Video models read body size, plumage, markings and posture; audio models denoise, slice, extract spectral features, discard wind and traffic, and match the result against the library.
Together they have identified more than 200 species in the province, including nationally protected ones, and produced records that would once have needed luck: in May, an infrared camera in Kunshan City Ecological Forest Park photographed a Malayan Night Heron — the first record for the city; in 2025, sound recorders at Hongze Lake picked up a White-throated Kingfisher, the species' first appearance in northern Jiangsu. Elliot's Pheasants were filmed in the wild, an Oriental Scops Owl's night song was logged repeatedly, and a Silver Pheasant was photographed leading chicks through forest at Longchi Mountain in Yixing — evidence that the species breeds there.
The three species the network records most often are worth knowing: the Light-vented Bulbul, the Vinous-throated Parrotbill and the Azure-winged Magpie, together more than 100,000 records.
What a machine still learns from a listener
Two lessons from these projects transfer straight to human ears. First, identification is a library problem: you cannot recognise what you have never heard, so a beginner needs a small, annotated set of sounds, not a shelf of them. Second, the same bird sounds different in different places and contexts, which is why the Jiangsu team built a local library rather than a generic one. A bird heard in Chongqing is not always the bird in the field guide.
None of this replaces a person who knows what they are hearing. It changes what that person is for: not counting birds one by one, but checking the machine, reading the trend, and noticing the record that no model was trained to expect — like a night heron in a city park.