The Albert Brand era
Albert R. Brand founded the Macaulay Library at Cornell in 1929. The recording technology of that era used phonograph cylinders, then disc recordings, with microphones connected to portable but heavy field equipment. Bird recording was a discipline practiced by a handful of dedicated specialists; the equipment cost meant that recordings were rare, carefully curated, and almost always made by trained ornithologists in good weather conditions. The recordings from this era represent a small archive of high-quality material captured under controlled circumstances. The species coverage is heavily biased toward common temperate-zone birds in accessible locations.
Magnetic tape recording, becoming practical in the 1940s and dominant by the 1950s, dramatically increased recording capacity and durability.
The magnetic tape era
Magnetic tape recording, becoming practical in the 1940s and dominant by the 1950s, dramatically increased recording capacity and durability. Field recordists like Peter Paul Kellogg at Cornell built the recording infrastructure that produced much of the foundational wildlife audio archive. Kellogg's recordings of Texas songbirds in the 1950s, his contributions to the Cornell library through the 1960s, and his teaching of younger generations of field recordists established the practices that defined wildlife audio for decades. The magnetic-tape era expanded the species coverage substantially but maintained the bias toward dedicated-recordist field expeditions to specific locations. Cassette tapes, easier and cheaper than reel-to-reel, expanded the practitioner base in the 1970s but the species and geographic biases remained substantial.
The digital revolution
Digital audio recording, becoming portable and affordable in the 1990s and 2000s, was the first major shift toward larger-scale recording. The combination of digital storage (no tape changes needed), better microphones at lower prices, and increasingly portable field-recording equipment (Marantz PMD recorders, then solid-state alternatives) reduced the barriers to wildlife recording. The species and geographic coverage in the Cornell archive grew faster in this period than in any previous era. But the recordings remained dominated by dedicated-recordist work, with citizen-science contribution still a small fraction of the total.
The phone-and-citizen-science era
Mobile phones with reasonable-quality microphones, the integration of audio with the eBird platform (early 2010s), and the broader citizen-science movement transformed the data-collection landscape. Suddenly any eBird user could attach an audio clip to a checklist as documentation, and the audio flowed into the Macaulay Library. The recording quality dropped (phone microphones are not field-recordist microphones) but the volume grew dramatically. By the mid-2010s, the majority of new Macaulay recordings came from eBird-affiliated contributors rather than from dedicated bioacoustics researchers. The species and geographic coverage continued to expand; the quality distribution became more variable.
The PAM and AudioMoth era
Autonomous recording units (ARUs) began transforming wildlife monitoring in the late 2010s. The AudioMoth, an open-source acoustic recorder developed by Open Acoustic Devices and released in 2016, is the iconic example: $80-100 per unit, runs on AA batteries for weeks, programmable for scheduled or trigger-based recording. The combination of cheap autonomous recorders, deep-learning audio classifiers (BirdNET[1], , NatureLM-audio[3]), and the broader citizen-science infrastructure has made acoustic monitoring of wildlife communities possible at spatial and temporal scales the field couldn't have approached just a decade ago. The data ecosystem is now characterized by very-large-volume citizen-science data plus growing-volume PAM-network data, with the older dedicated-recordist archive a small but very-high-quality subset of the total.
Why this history matters for AI
The training data for modern bioacoustic AI models reflects all of these eras. The high-quality dedicated-recordist recordings provide foundational reference material. The eBird-Macaulay era recordings provide species breadth and geographic coverage. The PAM-network recordings provide naturalistic continuous-audio context. The composite training set carries the biases of each era: under-representation of nocturnal and quiet species, geographic bias toward English-speaking-country contributors, temporal bias toward populated regions where citizen scientists live. AI models trained on this data inherit these biases. Recognizing where the training data came from is part of understanding what the models can reliably do and where they need additional development. CrowLingo's atlas, built on the smaller open-license Wikimedia Commons corpus rather than the larger but partially-proprietary Macaulay archive, makes a specific trade-off in this space — narrower data but cleaner licensing. The trade-off is real, deliberate, and a normal consequence of the history this audio infrastructure has had.