"The fridge sounds different lately." Everyone has said some version of this, and everyone has then discovered how useless it is as a report. Different how? Louder? Higher? A new rattle on top of the old hum? Human ears are superb at noticing that something changed and terrible at saying what. Memory of a sound fades in hours; a description of a sound barely survives one retelling.
The Spectrum tool exists for exactly this gap. Point the phone at a sustained sound — an appliance hum, a motor, a fan, an HVAC vent, an idling engine — record a few seconds, and get its frequency fingerprint: which frequencies are present and how strong each one is. That fingerprint is a thing you can save, file under the machine it came from, and compare against next month.
Why a spectrum beats "it sounds louder"
A microphone level meter answers one question: how much sound, total. But almost nothing that matters about a machine lives in the total. A washing machine's motor hum, its drum bearing, and the loose panel buzzing in sympathy all pour into the same loudness number — and a real change in one can hide behind a small change in another.
A spectrum pulls them apart. Rotating machinery is wonderfully honest in the frequency domain: a motor spinning at a steady speed produces a tone at its rotation rate and a family of harmonics — clean multiples of that fundamental — stacked above it. A fan adds a blade-passing tone at the rotation rate times the blade count. Mains-powered equipment carries hum at the line frequency and its multiples. Each mechanical process gets its own peaks, at its own frequencies, and the spectrum lays them out side by side.
That's what makes it diagnostic. "Louder" is one number moving. "A new peak appeared at three times the motor frequency" or "the peak that used to be narrow is now broad and surrounded by sidebands" points at a specific part doing a specific thing. Rattles show up as broadband energy smeared across many frequencies rather than a clean tone — audibly annoying, spectrally unmistakable, and clearly distinct from the hum underneath.
Recording it honestly
None of this works through the audio path phones use by default. That path is built for phone calls: automatic gain control that rewrites levels mid-recording, noise suppression trained to delete steady tones — a compressor hum is precisely what a call pipeline considers noise. Feed a spectrum analyzer from that path and you are analyzing the phone's opinion of the sound, not the sound.
So the Spectrum tool records through the platform's unprocessed path: raw samples, 48 kHz, mono, no gain riding, no enhancement. The steady hum the OS would have scrubbed away arrives intact, and the peaks in the spectrum are the machine's, not the pipeline's.
Why the fingerprint travels between phones: different phones hear the same machine at different loudness and with different tonal coloration, so the absolute height of a peak is partly a property of the microphone. But where the peaks sit, and how they relate to each other — the harmonic ladder, a new peak appearing where none was — are properties of the machine. The design rule across every tool is the same: build on relative, excitation-robust features, never absolute levels, so readings compare across devices without per-device calibration.
The measurement that matters is the second one
A single spectrum of an unfamiliar machine is interesting; a pair of spectra of the same machine is useful. Machines rarely fail loudly on day one — they drift. A bearing roughens, a mount loosens, a fan blade picks up an imbalance, and the fingerprint shifts weeks before the failure is obvious to the ear.
This is where the catalog earns its place. Every capture can be filed under an object — this fridge, that furnace blower — with repeated measurements accumulating under it over time. Record the machine while it's healthy; that's your baseline, and it costs ten seconds. When it "sounds different lately," record again under the same object and put the two fingerprints side by side. The question stops being "does anyone else remember what it used to sound like?" and becomes "which peaks moved?"
Consistency helps more than equipment does. Record from roughly the same spot, at roughly the same distance, with the machine in the same operating state. You're not chasing an absolute reading — you're keeping the comparison clean, so that a difference between Tuesday's spectrum and today's is a difference in the machine.
And when the story resolves — the technician finds the worn bearing, or the noise turns out to be a coin in the drum — the measurement can take a ground-truth label after the fact. A spectrum plus what it actually turned out to be is a labelled sample; enough of those, across many machines and many phones, is how sound-based diagnosis gets better at its job.
Hearing a saved spectrum again
One more thing the fingerprint enables, because it's a description of a sound rather than a recording of one: it can be turned back into sound. The viewer app can re-synthesize a saved spectrum's peaks audibly — "Play reconstruction" — so a measurement made last month can be listened to today, next to the machine, by the ear that will decide whether things have changed. It isn't the original recording; it's the fingerprint, resung. For "is this the same hum I measured before?", that's often exactly the right question to put back in front of a human ear.
Where it fits
Spectrum is the workhorse for sounds that sustain — hums, whines, drones, steady rattles. Sounds that happen once and die away, like a knock, want a different analysis: the tap test looks at how a brief ring decays rather than what a steady tone contains, and machines that shake more than they sing are better caught by the accelerometer, which hears the low-frequency motion a microphone misses.
All of it runs on-device — the analysis happens on the phone, offline, and the measurement is yours. For the wider picture of what phone sensors can measure, start with the tools overview.