Record ten to twenty seconds of the noise with a phone, drop the file above and the instrument reports the strongest narrow peak in the band you search: its frequency, its level, how far it stands above the background and how steady it stayed. Then change one thing, record again, and it tells you in plain words whether the peak moved, dropped or stayed where it was.
I built it for the point just before a return or a warranty claim. “It whines” gives a support agent nothing to check. A steady peak at one frequency in an uncapped menu, gone with a 60 FPS cap, same phone in the same spot, is a description someone can act on.
Record a capture the instrument can use
- Start the scene that makes the noise and leave it running. Mute the speakers.
- Put the phone outside the closed case, somewhere you can find again: on the desk a hand’s width from the side panel works. Keep it clear of fans and vents. Do not put a phone or anything else inside a running PC.
- Record ten to twenty seconds in the phone’s voice recorder. If the recorder offers an uncompressed or lossless setting, use it. The instrument reports where a recording stops carrying signal.
- Drop the file on the capture panel. That is capture A.
- Change one thing, leave the phone exactly where it was, record again and drop that file in as B.
The phone, the recording app and the position have to be the same for both captures. If they are not, the frequency comparison still holds and the level comparison does not.
What the readout means
- Peak frequency is the strongest narrow peak between the two edges of the search band. The capture is cut into overlapping frames, each frame gets a Hann window and a Fourier transform sized so one bin is at most 6 Hz wide (8,192 points at 48 kHz, which is 5.86 Hz), and the peak is placed between bins by three-point interpolation. A capture shorter than about a second gets a shorter transform and a coarser bin, and the readout prints the bin width it used.
- Level is in dBFS: decibels below the recording’s own digital full scale. It is useful only against another capture made the same way.
- Above background is how far the peak stands over the median of the spectrum around it. Anything under 10 dB is not listed, so a recording with no tone in it reports no peak instead of a guess.
- Steadiness has two parts: the share of the capture in which the peak was there, and how far its frequency ranged (5th to 95th percentile).
- The A/B result calls a frequency shift only when it is larger than both two analysis bins and the wobble measured inside the captures. It calls a level change at 3 dB or the frame-to-frame variation, whichever is larger. Smaller differences are reported as no change.
The search band starts at 1,000 Hz. That is a starting setting and not a statement about where coil noise sits. If what you hear is a buzz, lower the edge and look at the spectrum.
I test the arithmetic against synthetic tones with known answers. A tone in noise comes back within 1 Hz and 0.5 dB. A tone only 19 dB above the background comes back within 2 Hz and 1.5 dB. Those signals are test files, and none of them is published or presented as a recording.
What coil whine is
A coil is an inductor or a transformer. Würth Elektronik’s application note ANP118 gives two causes. A magnetic core changes its dimensions slightly in a magnetic field, an effect called magnetostriction, and when the field changes at an audible rate the core’s vibration reaches the air. The windings also feel a force from the field and vibrate. TDK’s guide to acoustic noise in power inductors lists the same two and adds attraction from the magnetized core. Both say a circuit board or a neighbouring part that resonates at the same frequency makes it louder.
The switching itself is too fast to hear. Würth puts the switching frequency of a buck converter at 75 kHz to 2.5 MHz in most cases, and TDK at several hundred kilohertz to several megahertz, against the 20 Hz to 20 kHz hearing range Würth gives. What reaches your ear is slower: a regulator that switches in bursts or skips pulses at light load, a switching frequency that falls as the load drops, or a load that itself changes at an audible rate. Each of those depends on what the hardware is doing, which is why the first steps of the sweep change the load.
Würth also notes that hearing is most sensitive between 2 and 8 kHz, so a whine there is noticed quickly, while one at 18 kHz goes unheard by most people in middle age. The range varies from person to person, which the note gives as the reason not everyone hears the same coil.
It is not always a coil. A 2022 review in Sensors describes ceramic capacitors that vibrate through piezoelectric and electrostrictive effects and drive the circuit board in the audible range. A spectrum cannot tell a coil from a capacitor. It can tell you the frequency, and whether it follows what you changed.
Why the frame cap comes first
It is the change a manufacturer names. Intel’s support article for Arc cards says the noise usually happens at a high frame rate and lists capping the frame rate and enabling V-Sync as its first two recommendations. AMD says its frame-rate target can reduce GPU power consumption and applies in menus and loading screens, where frame rates can run into the hundreds.
It does not work for everyone. The coil whine database asks each owner whether a cap changes the noise, and the answers on file are close to an even split. My own ASUS ROG Astral RTX 5080 OC is report number one there: a faint whine in uncapped menus that any cap removes.
Before you open a warranty claim
Read what the maker has already published. ASUS describes slight high-frequency noise from a graphics card as a characteristic of the components and says it will not affect the hardware. MSI describes the vibration as normal behavior and not a hardware defect. Intel says it should not damage the card or affect how it works, and tells owners with an extreme case to contact support. Corsair says power-supply coil whine is not exactly a defect and asks concerned owners to open a ticket. None of those pages decides an individual claim.
The copyable summary does not prove a defect. It shows that the noise is repeatable, where it sits and what changes it, in a few lines someone can read quickly. Send it with the exact model, the purchase date, the workload and a recording made from where you normally sit. If the return window is still open, decide between keeping, exchanging and testing another supply before it closes.
Stop and power down if the noise comes with smoke, a burning smell, visible connector damage or repeated shutdowns. That is a fault to have assessed, and the power supply failure guide covers it.
What a phone recording cannot tell you
How loud it is. The level is relative to the recording. For its voice-recognition input, Android’s compatibility document asks for a microphone response within ±3 dB only between 100 Hz and 4,000 Hz, strongly recommends no more than ±30 dB from 4,000 Hz to 22 kHz, and lets the sensitivity at 1 kHz sit 3 dB either side of its target. Inside that recommendation a peak at 9 kHz can read 30 dB higher or lower than the same sound would at 1 kHz, and two phones need not agree. No sound pressure figure comes out of this page.
Anything above half the sample rate. A capture sampled at 48,000 Hz can show nothing above 24,000 Hz; MDN’s description of browser frequency data uses the same example. The instrument prints that ceiling beside every result. Browsers resample audio when they decode it, so the page reads WAV files itself and, for other formats, decodes at the rate the file’s header states when it states one plainly.
Anything the recorder removed. If a recording carries no signal above some frequency, the instrument measures where it stops and says so, because a whine above that point would not appear at all.
Which part made the sound. A higher level beside the graphics card than beside the power supply is a clue about location. Locating the noise before swapping parts covers the rest.
Live microphone
Live listening appears only when the browser lets this page use a microphone. When it does, I ask for the raw signal with echo cancellation, noise suppression and automatic gain switched off, and the page prints what the browser granted, because a browser may keep them on. The capture goes through the same analysis as a file.
Reference recordings come later
There is no library to match against yet. Reference recordings from my own card, cooler and power supply come later. Until then the instrument works on your captures alone, and the database holds what other owners have reported by ear.
Sources
Checked November 17, 2026.
- Würth Elektronik, ANP118a: Acoustic noise & Coil whine effect, Levi Mark and Theo Ritzmann, May 9, 2025. Hearing range, magnetostriction, forces on windings, burst mode and pulse skipping, resonance, checking against load.
- TDK Product Center: Measures Against Acoustic Noise in Power Inductors. Causes of vibration, switching frequency against the audible range, load changes.
- Covaci and Gontean, “Singing” Multilayer Ceramic Capacitors and Mitigation Methods: A Review, Sensors 22(10), 3869, 2022.
- Intel support article 000093133, last reviewed December 3, 2024.
- AMD: Frame Rate Target Control.
- ASUS FAQ 1050014, updated March 16, 2023.
- MSI FAQ 10183, updated April 26, 2024.
- Corsair: What is coil whine?
- Android 15 Compatibility Definition, section 5.4.2. Microphone response and gain requirements for the voice-recognition source.
- MDN: AnalyserNode frequency data, decodeAudioData, Permissions-Policy: microphone.
- TechFuelHQ coil whine database, dataset 0.2.2.