How a cracked 1880s disc led AI to an 1817 poem

“not at the email he made” was an unpromising line to find in an 1880s recording. It appeared in a garbled Whisper transcription of a cracked wax-on-binder-board disc that UC Berkeley undergraduate Evan Odell was analysing. The recording sounded vaguely poetic, according to the lab’s August 2025 account. That impression suggested another route into the noise.[1]

Odell gave the transcript to ChatGPT and asked for nineteenth-century literature with similar metre and form. It suggested Charles Wolfe’s The Burial of Sir John Moore after Corunna. With the printed poem in hand, the researchers could recognise verses in the recording. Project leader Carl Haber called the identification “highly plausible”. The account does not identify the speaker.[1]

A record player made of measurements

Before a language model could offer that clue, the sound had to be recovered from an object too fragile for ordinary playback. IRENE, the optical preservation system used in this work, reads a recording’s physical surface without touching it. A Library of Congress account describes two ways to capture the grooves: a specialised microscope or a probe that measures surface height. The record rotates beneath the instrument as successive passes build up the image.[2]

The crucial next step is in software. Weaver follows the groove and calculates the motion that a stylus would have made, converting that virtual movement into an audio file. The same Library of Congress account describes an intern adding drawing-tablet support so an operator could trace difficult grooves by hand. A recoverable sound may depend on patient decisions about where a line continues.[2]

That approach grew out of particle-physics instrumentation. Berkeley Lab’s account of a separate restoration, published in 2013, describes mapping the surface in enough detail to process damage digitally. That project recovered Alexander Graham Bell’s voice from an April 1885 wax-on-binder-board disc. It establishes the technology’s history; it does not identify the voice in the later poem episode.[3]

A text waiting outside the recording

Wolfe’s poem was first published in 1817. Its opening is emphatic: “Not a drum was heard, not a funeral note”. The soldiers bury Moore at night, with bayonets standing in for proper digging tools and a lantern lighting the work. Its subject is a hurried burial; its shape is highly memorable. The University of Toronto’s annotated text supplies both the poem and its publication history.[4]

There is a small piece of evidence for how such a text travelled. In Poems Every Child Should Know, published in 1904, editor Mary E. Burt recalled encountering Wolfe’s poem in her childhood reading lessons. That recollection cannot tell us how this particular speaker knew it. It does show a route from a printed poem into the repertoire of texts people learned and spoke.[5]

The interesting move in Odell’s search was to change the question. Faced with an unreliable transcript, he asked for a work that might fit. A candidate poem supplies a sequence of words against which a listener can compare the old sound. That is a more tractable task than recovering every syllable without a starting point.

The printed text and the recording preserve different things. One gives the words in order; the other may preserve how someone once breathed, paused and spoke them. Bringing them together can make an obscure physical trace intelligible without replacing the performance that made it worth saving.

Report an error or suggest a correction