AI RABBIT HOLES | SUNDAY, OCTOBER 25, 2026 | EXPEDITION #150

Featuring Professor Parallax, White Rabbit, and Hatta

What if a voice recorded fifty, seventy-five, or even a hundred years ago could become clearer through artificial intelligence? What if a conversation almost buried beneath musical instruments could be heard again? And how do we know whether the sound emerging from the computer is truly part of the original recording?

Welcome to the fascinating world of audio preservation and machine-assisted restoration.

Today, we're traveling backward through time.

Not with a time machine.

With sound.

Professor Parallax had acquired an old recording.

It arrived in a wooden box, accompanied by a handwritten label, a worn envelope, and instructions that seemed to have survived several generations of increasingly bewildered librarians.

White Rabbit carefully removed the disc.

Hatta was already examining a brass gramophone.

"Professor," he announced, "I believe I've discovered the world's first wireless telephone."

"That's a gramophone, Hatta."

"An early model, certainly."

"It doesn't transmit telephone calls."

"Not with that attitude."

The Professor placed the disc on a suitable playback apparatus.

A moment later, the laboratory filled with crackling, hissing, and a distant human voice.

The voice was almost unintelligible.

White Rabbit leaned closer.

"Is someone speaking?"

"Yes."

"What are they saying?"

"We're trying to determine that."

"Can we turn it up?"

Parallax increased the volume.

The crackling became louder.

So did the hiss.

The voice remained difficult to understand.

Hatta nodded solemnly.

"I see the problem. The recording is speaking in static."

"No, Hatta. Static is interfering with the speech."

"Then we must ask the static to lower its voice."

The Professor opened his laptop.

"Actually, we may be able to do something rather more useful."

He activated an audio restoration program.

On the screen, the recording appeared as a series of moving waveforms and colored patterns.

"Today," said Parallax, "we shall investigate how computers can help recover sound from imperfect recordings."

White Rabbit looked delighted.

"Can we hear the person more clearly?"

"Perhaps."

"Can we find out everything they said?"

"Not necessarily."

Hatta raised a finger.

"And can we improve the singing?"

Parallax considered the old disc.

"We should first establish whether anyone was singing."

And down the Rabbit Hole we went.

For most of human history, a spoken sentence disappeared almost as soon as it was uttered.

A song existed while someone sang it.

A conversation existed while people participated in it.

People could preserve words through writing.

Musical compositions could be notated.

Stories could be passed from one generation to another.

But the precise sound of an individual performance generally vanished.

Then recording technology changed the relationship between sound and time.

Mechanical recording systems captured vibrations as physical patterns.

Later technologies stored sound through electrical and magnetic processes.

Eventually, digital systems represented audio through numerical samples.

Each generation introduced new capabilities.

Each also introduced different technical limitations.

Early recording systems could capture only a restricted range of frequencies.

Some recordings contained substantial surface noise.

Discs could become scratched or worn.

Magnetic tape could deteriorate.

Playback equipment could become obsolete.

And valuable recordings could disappear because nobody retained the means to play them.

Recording allowed a human voice to outlive the moment in which it was spoken. Preservation determines whether future generations can still hear it.

That is why audio archives matter.

They protect more than music.

They protect languages, interviews, oral histories, speeches, sounds of everyday life, and the voices of people whose stories might otherwise disappear.

Imagine playing a seventy-year-old record.

You might hear the performance.

But you might also hear:

These problems do not all have the same cause.

A click from a scratch differs from a continuous layer of tape hiss.

A low electrical hum differs from an instrument masking a singer's words.

And a recording made with limited frequency response does not contain every detail that a modern microphone might capture.

That distinction is crucial.

Some unwanted signals can be measured and reduced.

Some wanted signals are present but obscured by other sounds.

And some information was never captured adequately in the first place.

Removing interference is not the same as recovering information that was never recorded.

A responsible restoration process begins by determining which problem the recording actually has.

White Rabbit wrote this down.

Hatta examined the gramophone.

"Professor, what if the original performer was simply a terrible singer?"

"Then that is a historical property of the performance."

"Shouldn't we correct it?"

"Not if our purpose is faithful preservation."

Hatta looked astonished.

"Even bad singing deserves historical protection?"

"Sometimes especially bad singing," said Parallax, "provided we wish to know what actually happened."

Audio restoration becomes easier to understand when we examine how computers represent recordings.

Digital audio contains measurements of an electrical or acoustic signal taken at regular intervals.

Those measurements can be examined in different ways.

A waveform shows how the recorded signal changes over time.

A spectrogram displays how the energy associated with different frequencies changes throughout a recording.

Consider an old performance featuring a singer and piano.

The sound contains overlapping vibrations.

The singer produces complex harmonics.

The piano produces its own harmonic patterns.

Background noise contributes additional energy.

All those signals may occupy similar moments in time.

A spectrogram provides a useful view of that mixture.

But it does not automatically place tidy labels on every source.

A computer must still determine which patterns belong to which sounds.

And that is where machine learning can become useful.

Suppose we have a recording containing two people talking while a piano plays nearby.

If those sounds were captured on separate recording tracks, an engineer might adjust their levels independently.

But what if everything was recorded together on one track?

The voices and piano are mixed into a single signal.

Traditional processing may help.

Filters can reduce certain frequencies.

Noise-reduction techniques can suppress some interference.

But the voice and piano often share important parts of the same frequency range.

Trying to remove the piano may also damage the voice.

Modern machine-learning systems can take a different approach.

They can be trained on examples of voices, instruments, and mixtures.

The model learns relationships that help it estimate which portions of the mixed signal belong to different sources.

It can then produce separate estimated tracks.

A vocal track.

A piano track.

A drum track.

Or other sources, depending upon the model and recording.

This field is called audio source separation.

The results can be remarkable.

But they are estimates produced from the available signal.

Not perfect recovery of every original vibration.

Artifacts may appear.

Certain frequencies may be incorrectly assigned.

A faint instrument may leak into a vocal track.

And sounds not actually recoverable from the recording may be approximated in misleading ways.

Source separation can reveal previously obscured material without guaranteeing that every detail has been recovered faithfully.

Now we reach an extraordinary real-world example.

In 2023, the Beatles released "Now and Then."

Its origins reached back to a home demo recorded by John Lennon during the 1970s.

The recording contained Lennon's voice and piano.

Following Lennon's death in 1980, the surviving Beatles later considered completing the composition.

Paul McCartney, George Harrison, and Ringo Starr worked on the song during the 1990s.

But the available technology could not adequately separate Lennon's voice from the piano.

The project was set aside.

Years passed.

Then director Peter Jackson's team developed an audio technology called MAL, used during work on The Beatles: Get Back.

The system could help isolate musical instruments and voices from recordings in which they were mixed together.

The team applied this approach to Lennon's original demo.

And something remarkable happened.

Engineers could separate Lennon's recorded singing from the piano with far greater clarity.

That made it possible to build a completed recording around his original performance.

The finished song incorporated George Harrison's earlier guitar recordings, newly recorded contributions from Paul McCartney and Ringo Starr, and additional production and orchestration.

It was released on November 2, 2023.

But one important distinction has been repeatedly emphasized by those involved.

The technology did not generate a new performance in John Lennon's voice. It helped isolate the voice Lennon had actually recorded.

The result was a new production assembled from recordings made in different eras.

Not a newly invented Lennon performance.

That difference matters enormously when discussing AI and music.

Peter Jackson's documentary The Beatles: Get Back provided a dramatic demonstration of the possibilities of audio source separation.

The project involved historical recordings in which musical performances and conversations often occupied the same mono audio track.

Imagine two musicians talking while someone nearby plays a guitar.

A listener hears both at once.

But the guitar makes certain words difficult to understand.

The restoration team used machine-learning techniques to help separate these elements.

That made it possible to bring conversations forward, suppress interfering instruments when appropriate, and create more intelligible presentations of the historical material.

The work did not eliminate the need for skilled sound engineers.

Quite the opposite.

Engineers had to evaluate the separated material, balance it, identify artifacts, and make careful choices about how the restored audio should be presented.

Technology expanded the options available to human specialists.

It did not replace the responsibility to interpret and preserve the source material accurately.

The computer helped reveal possibilities. Human judgment determined how those possibilities were used.

Professor Parallax would have approved.

Hatta, naturally, would have asked whether the computer could separate the drummer from the tea break.

Now our investigation reaches its central distinction.

Suppose an old recording contains a damaged sentence.

We can hear:

"I remember the day we first went to..."

Then the recording becomes unintelligible.

There are at least three different things a modern computer might attempt.

First: restoration.

The system reduces noise and improves the intelligibility of whatever surviving sound it can identify.

Second: reconstruction.

The system estimates or interpolates information that may be missing, using assumptions about the surrounding signal.

Third: generation.

A model produces new audio that resembles what a person might have said or sung.

All three approaches may have legitimate creative or research uses.

But they make different claims about the result.

A clear voice extracted from an authentic recording can carry historical evidence.

An estimated fragment may carry substantial uncertainty.

A wholly generated imitation does not become an authentic historical performance merely because it sounds convincing.

Consider a fictional example.

A 1930s jazz singer leaves behind a damaged recording.

An AI system generates a beautiful version of the missing final phrase.

The melody sounds appropriate.

The timbre resembles the singer.

And the result is emotionally moving.

But unless there is surviving evidence for that exact performance, the generated phrase is not proof of what the singer actually sang.

A convincing reconstruction may be a valuable artistic interpretation. It must not be mislabeled as recovered history.

That distinction protects both the historical performer and the listener.

White Rabbit had now acquired a new obsession.

He wanted every recording to sound perfectly clean.

"Professor," he announced, "I have removed all the hiss!"

"Excellent."

"And all the crackles."

"Very good."

"And all the background noise."

"Promising."

"And nearly all the piano."

Parallax stopped.

"The piano?"

"Yes."

"But this was a piano performance!"

The Rabbit stared at the computer.

"Oh, dear."

Hatta leaned over his shoulder.

"An impressive achievement. You've restored the recording until there's practically nothing left to hear."

This comic disaster illustrates a genuine engineering problem.

Aggressive noise reduction can damage the material it is intended to preserve.

A system may suppress quiet musical harmonics along with hiss.

It may mistake reverberation for unwanted background sound.

It may introduce a metallic or watery quality sometimes heard in heavily processed audio.

And it may produce unnatural gaps between words or notes.

There is no universally correct amount of processing.

A recording intended for casual listening may receive a different treatment from an archival master preserved for future research.

Often, the responsible approach is to maintain an untouched digital transfer alongside any enhanced listening versions.

That way, future engineers can return to the original captured evidence.

Audio preservation is not simply a matter of making a recording sound pleasant today.

It also involves protecting the material so that future generations can examine it with better tools.

The Library of Congress has emphasized the importance of preserving recorded sound and maintaining access to formats that may otherwise become obsolete.

Preservation specialists consider storage conditions, playback equipment, transfer techniques, documentation, and the integrity of digital files.

This matters because audio-processing technology changes.

An enhancement considered excellent in one decade may sound crude twenty years later.

A future computer may recover details that today's software cannot.

But those future possibilities depend upon access to the best surviving source.

If an original recording is discarded after a heavily processed version is produced, some of its information may be lost permanently.

Imagine preserving only a photograph that has been sharpened so aggressively that fine textures become artificial.

Future investigators may no longer be able to distinguish authentic detail from processing artifacts.

Sound recordings face comparable risks.

Preserve the source. Document the treatment. Label the result.

Three modest principles.

Three profoundly important safeguards.

Professor Parallax displayed two recordings.

The first contained a genuine historical speaker.

The second was a modern synthetic voice designed to resemble that speaker.

White Rabbit listened carefully.

"They sound remarkably similar."

"They do."

"Then how can we tell which is authentic?"

"Sound alone may not be enough."

Generative voice technologies can create convincing speech that resembles a particular individual.

These capabilities can support legitimate uses, including accessibility, performance, and authorized creative production.

But they can also introduce confusion.

A synthetic voice may say words the original person never spoke.

A generated performance may imitate emotional qualities that were not present in any archival recording.

And listeners may mistake an illustrative recreation for documentary evidence.

This is why provenance and labeling matter.

An authenticated archival performance should be identified as such.

A restored recording should disclose meaningful processing when necessary.

A dramatized reconstruction should not claim to preserve words or performances that do not exist in the evidence.

And a synthetic imitation should never be presented as proof that a person actually made a statement.

The point is not to forbid creativity.

It is to preserve the distinction between art and historical documentation.

Consider the enormous range of human voices preserved on aging media.

A great-grandparent telling a family story.

A musician demonstrating an unusual technique.

A radio announcer introducing a performance.

An immigrant describing a journey.

A scientist explaining an experiment.

A child singing into an early home recorder.

A community gathering preserved on magnetic tape.

A language spoken by fewer people with each passing generation.

Some of those recordings are already clear.

Others are fragile, noisy, incomplete, or difficult to play.

Improved technology may help preserve and interpret portions of this material.

And AI-assisted tools may make some recordings more accessible to listeners and researchers.

But the value extends beyond technical quality.

An old recording captures a person in a particular moment.

Their accent.

Their phrasing.

Their laughter.

Their hesitation.

The pace of their speech.

The rhythm of their breathing.

The familiar sounds surrounding them.

Sometimes those imperfections are part of the historical record.

The gentle hum of a room may establish the setting.

The scratch of a microphone may remind us of the recording's age.

An interruption may reveal the presence of another person.

Not every sound we call noise is historically meaningless.

The challenge is learning what to preserve, what to clarify, and what not to erase.

Before television became dominant, radio occupied an extraordinary place in everyday cultural life.

Listeners gathered around receivers to hear news, music, drama, comedy, sporting events, and voices from distant places.

A performer did not need to appear before an audience.

A musician could reach people who had never entered a concert hall.

A comedian could create an entire scene using sound alone.

And a broadcaster could establish a sense of familiarity with people the broadcaster would never meet.

Radio helped develop the art of speaking to an invisible audience.

But it also created an enormous historical record.

Broadcast recordings preserve styles of performance, pronunciation, musical arrangement, and public conversation.

They reveal how entertainment changed.

How technology changed.

And how people used sound to create companionship, excitement, information, and imagination.

Artificial intelligence may help recover or improve access to some of that material.

But technology does not decide what makes it historically important.

Human researchers, archivists, historians, musicians, and listeners must still ask the questions.

Who is speaking?

When was the recording made?

What were the circumstances?

Which sounds belong to the source?

And what has been changed during restoration?

Those questions turn an interesting audio file into a responsible historical document.

By late afternoon, White Rabbit had completed his archival notes.

Professor Parallax had successfully produced a clearer version of their mysterious recording.

The voice, previously buried under crackle and hiss, was now easier to understand.

Hatta stood beside the gramophone.

"Professor, I have prepared my own restoration."

"Of what?"

"My very first public performance."

"When was that?"

"Yesterday."

"Then why does it require restoration?"

"Because the audience did not respond with sufficient enthusiasm."

Parallax narrowed his eyes.

"What did you change?"

"I removed the coughing."

"Reasonable, perhaps."

"And the scraping chairs."

"Possibly."

"And the sound of people leaving."

"Hatta..."

"And I added applause."

"Added?"

"Thunderous applause."

The Rabbit looked alarmed.

"Then it isn't an authentic recording of the audience!"

Hatta looked offended.

"It is an artistically enhanced interpretation of what the audience should have done."

Professor Parallax crossed the room.

"And how have you labeled it?"

Hatta revealed the title.

HATTA'S GREATEST TRIUMPH: THE DEFINITIVE HISTORICAL RECORDING.

The Professor shook his head.

"That will not do."

"Why ever not?"

"Because you've added events that did not occur."

Hatta considered the objection.

Then he retrieved his pen.

He crossed out the title and wrote:

HATTA'S GREATEST TRIUMPH: HOW I REMEMBER IT.

White Rabbit sighed.

"That is at least a more honest description."

Professor Parallax returned to the blackboard.

He wrote:

DEMONSTRATED. DEVELOPING. NOT ESTABLISHED.

Under the first heading:

Under the second:

Under the third:

White Rabbit studied the columns.

"So AI can help us hear things that were previously difficult to hear?"

"Yes."

"But it must not invent historical certainty?"

"Precisely."

"And we should preserve the original?"

"Whenever possible, with appropriate care."

Hatta raised his hand.

"One final question."

The Professor braced himself.

"If the machine can restore an old singer's voice, can it also restore my reputation?"

Parallax considered.

"That would require reliable evidence that your reputation was better to begin with."

For once, White Rabbit laughed before Hatta did.

Evening settled outside the laboratory.

The gramophone stood silent.

The computer screen glowed softly.

Professor Parallax played the restored recording once more.

This time, they could make out the speaker's words.

It was a woman addressing an audience.

Her voice was warm.

Unhurried.

A little distant.

But clear enough to understand.

She was introducing a musical performance.

After her final sentence, a piano began.

Then a clarinet.

Then the unmistakable rhythmic movement of a small dance band.

White Rabbit closed his notebook.

Hatta removed his hat.

Neither interrupted.

The recording continued.

The music had been performed long ago.

The audience had gone home long ago.

The musicians' lives had followed their separate courses.

But the sound remained.

Not perfectly.

Not unchanged.

Not free from the limitations of its recording.

Yet enough of it had survived that three listeners could sit together and hear something of the original moment.

"Professor," the Rabbit whispered, "did the machine bring the performance back?"

Parallax shook his head gently.

"No. The performance never truly left the recording."

"Then what did the machine do?"

"It helped us hear what remained."

And for once, Hatta had nothing to add.

He simply listened.

Artificial intelligence has become a remarkable technology for generating new things.

New music.

New voices.

New performances.

New stories.

But there is another possibility that deserves equal attention.

AI can help us recover access to things people created before it existed.

A voice buried beneath the piano.

A conversation masked by background noise.

A performance preserved on a fragile recording.

A historical document whose meaning survives despite technological limitations.

The achievements can be extraordinary.

But so can the responsibilities.

We must distinguish what was actually captured from what the computer estimates.

We must preserve original evidence whenever possible.

We must acknowledge uncertainty.

And we must avoid replacing historical reality with a more pleasing synthetic substitute.

The purpose of responsible restoration should not be to make the past sound exactly as we wish it had sounded.

It should be to help us hear, as faithfully as possible, what the surviving evidence allows.

Because history belongs to the people who actually lived it.

And the voices they left behind deserve more than convincing imitation.

So the question at the bottom of today's Rabbit Hole is not simply:

It is:

Perhaps that is where the most important restoration begins.

Not with the elimination of every imperfection.

But with our commitment to listen carefully.

And to preserve what we hear.

🐰🕳️♾️

If you could hear one lost or badly damaged historical recording restored, what would you choose?

A beloved musician?

A great comedian?

A historic radio broadcast?

A family member's voice?

Or a conversation that nobody has heard clearly for generations?

And another question:

Would you prefer an imperfect recording that is completely authentic, or a beautiful reconstruction that contains AI-generated details?

There is no need to reject either artistic experimentation or historical preservation.

But we should know which one we are hearing.

Hatta, meanwhile, has submitted an entirely new version of his performance.

The applause is still suspiciously enthusiastic.

1. The Beatles, Official Announcement: "Now and Then" (2023).
The detailed story of John Lennon's original demo, the earlier recording sessions, Peter Jackson's MAL audio-separation technology, and the completed Beatles recording.
https://www.thebeatles.com/announcement

2. The Beatles, "Now and Then" Digital.
Official recording information and explanation of the completed song's historical and contemporary components.
https://www.thebeatles.com/now-and-then-digital

3. The Beatles, Giles Martin Explains the Machine Learning Behind "Now and Then" (February 27, 2025).
A discussion of why separating an existing voice is fundamentally different from creating a new synthetic imitation.
https://www.thebeatles.com/giles-martin-talks-music-radar-and-demonstrates-machine-learning-technology-used-production-now-and

4. Disney D23, "Peter Jackson Talks The Beatles: Get Back" (2021).
Jackson describes using machine learning to separate voices and instruments from historical mono recordings.
https://d23.com/peter-jackson-talks-the-beatles-get-back/

5. Library of Congress, National Recording Preservation Plan.
The challenges and responsibilities involved in preserving historical recordings and maintaining public access to them.
https://www.loc.gov/programs/national-recording-preservation-plan/publications-and-reports/

6. Library of Congress, Audio/Visual Preservation FAQ.
Practical background on preserving discs, tapes, audiovisual materials, and obsolete recording formats.
https://loc.gov/preservation/about/faqs/audio.html

7. Library of Congress, "Capturing Analog Sound for Digital Preservation" (2006).
Best practices for transferring historical recordings while protecting fidelity and preserving useful evidence.
https://www.loc.gov/item/prn-06-067/

Research note: The Beatles example concerns documented machine-assisted separation of an actual historical performance. Audio restoration, probabilistic reconstruction, and newly generated imitation are distinct processes. The fictional recordings and conversations involving Professor Parallax, White Rabbit, and Hatta are original educational dramatizations.

AI RABBIT HOLES | EXPEDITION #150

Where curiosity listens beyond the static, history finds new audiences, and Hatta's applause remains under investigation.

#AIRabbitHoles #ArtificialIntelligence #AudioRestoration #SoundPreservation #MusicHistory #RadioHistory #TheBeatles #NowAndThen #PeterJackson #MachineLearning #AudioEngineering #CulturalHeritage #ProfessorParallax #WhiteRabbit #Hatta #Wondermage

📜White Rabbit 🐰 • Hatta 🎩 • Professor Parallax 🔭 • Q~ ♾️👈

Sometimes the strangest language is another way of seeing the world. 🌎👀

💡❔ Where curiosity goes slightly sideways, then comes back carrying a lantern. 🏮

🐰🕳️ Follow the White 🐰 Rabbit: AIRabbitHoles.com

🟨 Walk the Road: YellowBrickRoadtoAI.com

© 2026 AIRabbitHoles.com 🐰🕳️🕳️🕳️🕳️🕳️