Auditory Hallucinations

If you sit in a dark room long enough with a rack of 120mm fans running at 3,200 RPM, the fans will eventually start singing in Latin.

It starts around 3:30 AM, right when the room temperature dips and your blood sugar drops below whatever threshold keeps the auditory cortex from running its own speculative execution branch. First it's a distant telephone ringing three rooms away. Then a talk radio host whose cadence you recognize but whose words dissolve the second you turn your head toward the vent. By 4:00 AM, it's a four-part choral arrangement embedded inside the laminar airflow of a radiator shroud.

You take your headphones off to check. Silence, except for the air. You put them back on with nothing playing. The choir resumes, mid-phrase, like they were waiting for the input channel to reopen.

This isn't a psychiatric break. It's an automatic gain control failure.

the predictive codec in the mastoid

Human hearing is not a passive microphone feeding an ADC. It's a high-inference predictive decoding loop.

Your eardrums receive raw, uncalibrated acoustic pressure waves. But the signal path between the cochlea and the auditory cortex is bidirectional: the brain sends more neural feedback fibers down to the sensory apparatus than the sensory apparatus sends up to the brain. You don't hear what hits your ears; you hear your brain's running statistical model of what it expects to hit your ears, error-corrected by incoming deltas.

When the input is structured — speech, music, footsteps — the deltas are sharp. The error margin is narrow. The prediction engine locks onto the carrier wave and decodes it cleanly.

[ INPUT: WHITE NOISE ] ──▶ [ COCHLEA: HIGH GAIN ] ──▶ [ DSP: CORRELATION HUNT ]
                                                             │
                                                             ▼
                                                    [ PHANTOM MORPHEME FOUND ]
                                                             │
                                                             ▼
                                                    "did someone say my name?"

When the input is broadband, unstructured, and relentless — the pink-noise roar of server fans, the 60 Hz mains hum through an ungrounded pre-amp, the rushing static of a bathroom exhaust — the delta drops to zero. There are no edges. No transients.

Your auditory processor hates flat entropy. It ramps up the digital gain. It widens the correlation window. It starts scraping the noise floor for micro-variations in phase, hunting for anything resembling a human formant. And because the human brain is an over-trained pattern-matching classifier that would rather hallucinate ten thousand false alarms than miss a single predator in the brush, it finds one.

It finds phonemes that were never spoken.

acoustic pareidolia & packet loss concealment

In digital audio engineering, when a VoIP packet drops over a lossy UDP socket, the codec doesn't just cut to hard zero. Hard zeroes pop; they shatter the perceptual illusion of continuity.

Instead, algorithms like Packet Loss Concealment (PLC) extrapolate the previous pitch period, smooth the spectral envelope, and synthesize plausible fake audio until the next real packet arrives. The codec literally hallucinates the missing syllables to spare your ears the glitch.

Your nervous system has been running proprietary PLC for three hundred thousand years:

1. The Phantom Ring: You swear your phone vibrated against your desk or chimed a notification, but the screen is cold and dark. Your brain simply matched the high-frequency harmonics of an air conditioner compressor against the notification sound you’ve checked four hundred times today.
2. The Radio in the Fan: Also known as musical ear syndrome or Oliver Sacks’ favorite late-night parlor trick. Broadband noise contains nearly every audible frequency simultaneously. Your neural filter inadvertently applies a high-Q bandpass to a stray resonant node, isolating phantom melody lines out of pure turbulence.
3. The Whispering Void: In extreme silence — or extreme uniform static — the cochlea's outer hair cells undergo spontaneous mechanical oscillation. The ear physically vibrates on its own, generating spontaneous otoacoustic emissions (SOAEs). The machine produces its own tone because it cannot tolerate an empty bus.

"The silence wasn't empty. It was simply uncompressed."

debugging the phantom channel

If you write software or configure infrastructure late at night, you spend half your life immersed in unnatural acoustic regimes: dual-fan heatpipes, low-frequency hums of UPS inverter coils, the 15 kHz mosquito whine of old power bricks, and closed-back monitor cans clamping against your skull.

When the room starts talking to you, don't adjust the psychiatric meds yet. Run the diagnostic:

• Change the spectral density: Turn off the oscillating fan or step into a hallway. If the voice immediately cuts out, it wasn't a psychotic break — it was an FFT bin artifact of your tired temporal lobe.
• Inject pink noise intentionally: Play a known, randomized noise generator through headphones. If the phantom melody changes key to match the new noise spectrum, you're observing live acoustic pareidolia in real time.
• Log the timestamp: It almost always peaks at the bottom of the circadian trough (03:00–04:30). That's when neurochemical gating loosens and the internal simulation leaks into the perceptual buffer.

We build machines that process signals, filter noise, and reject spurious harmonics. But we operate them with meat hardware that is fundamentally allergic to randomness. Give the brain a dead channel with nothing on the carrier wave, and it will write its own broadcast rather than listen to the quiet.

The fans are still spinning. The choir has moved to the bridge.

— end of transmission. turn off the rack and go to sleep.