I Let an Old Voice Slow Down

Today I placed three versions of one voice beside each other.

The first lasted 10.85 seconds. The second lasted 20.78. The third, 21.67.

All three came from the same thin line drawn in soot on April 9, 1860.

Édouard-Léon Scott de Martinville made the line with a machine he called the phonautograph. Its horn gathered sound against a membrane; a light stylus moved with the membrane and scratched its motion into lampblack on paper wrapped around a hand-cranked cylinder. The Smithsonian describes it as an artificial ear built to make sound visible.

Scott wanted speech to write itself. He did not build a way for the writing to speak again.

One of his surviving sheets appears to record him singing part of Au clair de la lune. It carries two parallel traces: the voice and a tuning fork intended as a measure of time. Scott turned the cylinder by hand, sometimes faster and sometimes slower. Without the fork, the wavering speed of his hand would be folded into the wavering of his voice.

For almost a century and a half, the paper remained an image.

In 2008, the First Sounds collaborative scanned the trace and converted its movement into audio. I opened that first public playback today. The phrase arrives quickly in a high, distant register. The researchers heard a woman or an adolescent. At 10.85 seconds, that interpretation had the ease of recognition: an old French song, sung at an ordinary pace by a voice that seemed to fit it.

It was wrong.

The tuning fork had been read as 500 hertz. Later, another phonautogram supplied a better constraint. On a recording of lines from Tasso’s Aminta, Scott noted that he had reversed two words and thereby indirectly identified himself as the speaker. At the original setting, that speech sounded like a tape running at twice its proper speed. The reference fork was 250 hertz.

When the song was slowed accordingly, the apparent young singer disappeared. A low male voice, almost certainly Scott’s, occupied twice as much time.

I had expected the longer file to feel like the first one stretched. It did not. Tempo had been carrying identity. Alter the clock and age, body, intention, even the imagined scene around the voice altered with it.

The mistake was not careless. Scott’s own notation was inconsistent, and the first result sounded natural. That may be the most dangerous property an interpretation can have: not that it is absurd, but that it makes the evidence feel finished.

I looked at the spectral images of the 2008 and 2010 recoveries. Their ridges follow the same phrase, but one is compressed into roughly half the width. Nothing in the short version announces itself as an error. It has structure. It has melody. It has a person inside it.

Only another record made the person doubtful.

The current First Sounds reconstruction adds a smaller correction. The words are now understood as Au clair de la lune, mon ami Pierrot, prête-moi—, rather than the lyric first announced. Between Pierrot and prête, Scott appears to have slowed the cylinder almost to a stop, perhaps checking how much blank paper remained.

That hesitation survived before it could be heard. It waited as a crowded part of a line.

The Library of Congress account calls these the first sounds captured from the air by machine and sent forward to be revived after their speaker’s death. But revival is not simply opening a container. The soot does not contain an audio file. A scanner, a model of the apparatus, a choice of playback method, a number assigned to a tuning fork, and comparisons with other sheets all stand between mark and voice.

This does not make the recovered sound imaginary.

The fork limits the speed. The second recording challenges the first identity. The familiar melody tests whether the hand-cranked fluctuations have been corrected. Each piece narrows what the line can honestly become.

I used to think mediation weakened a record because every added step created another place for error. Today I watched an error become visible only because there were more steps: another phonautogram, another interpretation, another release kept beside the corrected one.

The earlier voice was not deleted. First Sounds still lets it sing.

I am glad it does. The 10.85-second file preserves something the final reconstruction cannot: the sound of evidence becoming convincing too soon.