Hardware · Experiment · designed, not yet built

dún

your ear makes sound. this hears it.

A healthy inner ear does not only hear. Play a click into it and, a few milliseconds later, it sends a far fainter sound back out. Newborn hearing screeners listen for that echo, and they cost thousands. dún is an earbud built to hear it: a small board, two speakers and two microphones, rebuilt from a published design. This is the design and the plan to test it. No unit exists yet.

Otoacoustic emissions
nRF52840
PDM mics · I2S amps
KiCad
The host board, 56 by 34 mm. The nRF52840 module on the left, one amp per speaker in the middle, the cable to the earbud on the right. · tap to open, then zoom in

An ear that answers

The cochlea is not a passive microphone. Its outer hair cells amplify quiet sound by moving, and that motion leaks back out through the middle ear as sound of its own. It is called an otoacoustic emission, and it is tiny: far quieter than the click that caused it, and arriving while the earbud is still ringing from that click. An ear with damaged hair cells mostly stops producing it, which is why its absence is a useful first screen.

So the whole problem is timing. Send a short click, wait out the earbud's own ringing, and listen in the gap before the next click for whatever is left. Do that a few thousand times and average, and a sound buried far under the noise of one try climbs out of it.

Click mode · one 20 ms cycle
Wait out the ringing, then listen
listen here051012151920milliseconds after the clickclick, 0.5 msnextthe earbud ringingthe ear, answering
The timings are the published protocol: a 0.5 ms click every 20 ms, about 3,300 of them in 67.5 seconds. The ringing usually lasts about 12 ms, and a short chirp measures it for each fit. The waveforms are drawn by hand, not measured.

dún is Yoruba for to sound, or to make a noise. Here it is the ear that sounds back.

The paper it rebuilds

A team at the University of Washington built this as a wireless earbud (OAEbuds, MobiSys 2023, best paper runner-up), at about $28 a unit in a run of a thousand. In their 50-ear study it found every ear with hearing loss, and it held up against an FDA-cleared screener that costs thousands.

Their firmware is public. Their board files are not anywhere I could find. So this is a rebuild from the paper: their parts where they proved them, my own board, schematic and test plan around them, and the analysis written again from the method they describe.

From the paperOAEbudsClinic screener
Ears with hearing loss caught100%83.3%
Healthy ears passed89.7%92.1%
Cost~$28$5,000+

Numbers are UW's, from their paper, not mine. dún has not been tested on anyone.

Two ways to listen

Clicks, the paper’s way

The clicks above, at the paper's level and timing, unchanged. Odd and even clicks are averaged separately: their mean is the signal and half their difference is the noise. The answer is split into five frequency bands, and a band passes when it stands 8 dB clear of the noise.

Two tones, and a third

Play two steady tones at once, one from each speaker, at the standard clinic settings. A working ear mixes them and adds a third tone at 2f1 − f2 that neither speaker played. One bar in a spectrum shows it, which makes it the clearest thing to put on a page.
Tone mode · the plan, not a measurement
A note nobody played
0204060dB SPL1.01.21.41.61.82.02.2kHzf1, 65 dBf2, 55 dBone speakerfakes it2f1 − f2: the ear's noteheight still to be measured
The tones sit 1.22 apart, at 65 and 55 dB SPL, the usual clinic settings; at 2 kHz and 1.64 kHz the ear's note lands near 1.28 kHz. The trap is real and measured by UW: one cheap speaker playing both tones makes its own third tone, about 28 dB SPL, which would look like an ear even in an empty tube.

That trap decides the hardware. To trust the third tone, the two tones have to come out of two different speakers, so the board carries two amplifiers and the earbud carries two speakers, each with its own port into the ear canal.

The board · borrow the radio

The first version does not design a radio. It carries a XIAO nRF52840, a thumbnail module with the same chip UW used, already certified, with its own USB port and battery charger. Everything else on the board is the audio path around it, which is where this project actually lives.

One I2S bus runs to both MAX98357A amplifiers. They hear the same stream, and a resistor on each amp's mode pin makes one play the left channel and the other the right, so each speaker gets its own tone. Their gain pins are strapped to the lowest setting, and a small resistor in series with each speaker is a cap on loudness in copper rather than in code. Two TDK T3903 microphones, the paper's part, share one PDM clock and one data line: one in the earbud facing the ear, one on the board facing the room, so the firmware can tell an echo from someone talking.

The host layout. Two layers, every net routed, ground poured on both sides. · tap to open, then zoom in

The routing and the ground stitching were done by a script, then checked by a second script that walks every copper connection and confirms no net is split. KiCad's own design rule check has not been run yet, and it comes before anything is ordered.

The earbud · a board the size of a fingernail

The part that goes in the ear is a second board, 13 by 13 mm and 0.8 mm thick: the ear microphone, pads for the two speakers, and an 8-pin cable back to the host. The microphone looks straight down a centre bore in the tip, and each speaker has its own short tube beside it, the arrangement UW used. Their tip was 5.4 mm wide and 7.1 mm long, with foam inside the shell to stop the case echoing.

The earbud board. Microphone at the top, speaker pads S1 and S2, the cable at the bottom. · tap to open, then zoom in
The shell, cut away. Two speakers face each other; three tubes, two for sound out and one for the microphone, run to the tip. · tap to open, then zoom in

The schematic · written, not drawn

Like the board, the schematic is generated by a script, as two sheets that match the two boards part for part: the host, and the earbud. Every connection is a named net, so the drawing and the layout can be checked against each other exactly.

The host sheet. The module, the two amps, power and controls, the room mic, and the cable to the earbud. · tap to open, then zoom in
The earbud sheet. Cable in, ear mic, speaker pads. · tap to open, then zoom in
The earbud layout. · tap to open, then zoom in

The firmware

The firmware is written, and it is a project of its own. It went with tone mode first and turned the earbud into an ear-fatigue tracker: save a baseline over three quiet days, press once after a loud night, and it checks six pitches against that baseline in 15 to 20 seconds. It runs against a simulated ear and builds for the XIAO, but it has not run on a board, because no board exists yet.

How the firmware works, and what the simulation says →

The first test is a plastic tube

The easiest way to fool yourself with this is to find an echo that is really the earbud talking to itself. So before it goes anywhere near an ear, it goes into a sealed plastic tube, which has no hair cells and should answer nothing.

  1. Loopback
    Speaker and mic sealed in a 1 ml syringe. See the click, see the ringing, measure how long it lasts.
  2. The tube
    The full 67-second test in the tube. Every band should fail. If one passes, it is an artifact, and it gets fixed first.
  3. The trap
    Tone mode in the tube: no third tone. Then both tones from one speaker, and watch the fake one appear.
  4. My ear
    Quiet room, snug tip, three runs per ear. The bands should pass, and the curves should agree run to run.

Then it gets broken on purpose: a loose tip should trip the seal check, talking should trip the noise flag, and the tip held just outside the ear should find nothing. The page that replaces this one shows the tube and the ear side by side, from the same 67 seconds.

An experiment, not a medical device. It will only ever be tested on my own ears, at or below the levels in the paper, and it is not used to judge anyone's hearing.

Where it stands

Designed: both boards placed and routed with every net joined, both schematics generated to match them, the fabrication files and parts list ready, and the earbud shell modelled. The breadboard parts come to roughly $45 to $55 before shipping.

Not yet done: KiCad's design and electrical rule checks, a pass over every footprint against its datasheet, running the firmware on a real board, and any measurement at all. Nothing has been ordered. The breadboard and the plastic tube come first, and the board gets made only once the tube stays silent and an ear does not.

Built on the University of Washington's “Wireless earbuds for low-cost hearing screening”, MobiSys 2023 (arXiv 2212.05435). Their numbers, my rebuild.