Leo is a pair of glasses for people with cochlear implants. In a loud room, it raises the one voice you are trying to follow above the noise around it. For Deaf signers it works the other way, and speaks their signing aloud to people who don't sign.
To hear which direction a voice comes from, you need microphones spread across the head. Glasses already sit there. Here is where each part goes on the DK-1 frame.
Drag to rotate. The frame is generated from the DK-1 design targets and rendered in your browser: titanium rims and temples, prescription-compatible lenses, six microphone ports across the brow bar and temples for aperture, two sign cameras under the bridge for the outward channel, and the waveguide on the right lens. Compute and radio are split across the temples for thermal and mass balance. To be explicit: this frame has not been built. It is the layout our software is being written against, not a photograph of something on a desk.
AI glasses for cochlear implant recipients.
The software runs. The frame does not exist yet.
A whole Leo, on a face, working.
You get left out when you can't pull one voice out of the noise, and when the room can't understand the voice you have. For decades the first half was a hardware problem with no good fix. What changed is recent: AI can now pull overlapping voices apart in real time and label who said what, on a chip small enough to wear [15][16][17]. Leo puts that on your face, and pairs it with a channel that runs the other way.
Six microphones and an on-frame model split the room into separate voices in real time, lock onto the person you're attending to, and caption each line tagged to who said it, so a table of people talking over each other becomes one thread you can actually follow.
For a cochlear-implant wearer, whose hearing plateaus at a handful of usable channels, that ratio is the only lever left. How the acoustic stack works →
Two cameras under the bridge watch the wearer's signing space; an on-device model reads the hands and the frame speaks the words aloud. A Deaf signer stops needing a phone, a notepad or an interpreter just to be understood by someone who doesn't sign.
Everything is read on-frame and the video is discarded; the wearer picks the voice and confirms before anything is spoken. How sign → speech works →
Every other company building hearing AI is competing for space inside the ear. For the people we build for, that space is gone, and it was gone before anyone shipped an earbud. What is left is the face, and it turns out to be the surface our users were most likely to be wearing something on anyway.
A cochlear implant recipient wears a sound processor on the ear and against the side of the head, every waking hour. That is not a preference and it is not coming off. Any product shaped like an earbud is structurally locked out of this user, not because the software is weak but because the geometry is taken.
Glasses are the only unoccupied surface on the head, and they are the only one with the sightlines this problem needs. An array on the temples has aperture the ear cannot offer, the bridge looks where the wearer looks, and the same mount that hears the talker can see the talker's mouth and the signing space.
Vision problems affect somewhere between 44 and 65 percent of deaf and hard-of-hearing people, against 17 to 30 percent of normal-hearing people [26]. Roughly double, and it is not a coincidence.
The retina and the cochlea differentiate from the same layer in weeks six and seven of gestation [26]. What damages one frequently damages the other, whether genetic, hypoxic or viral. Usher syndrome, the leading inherited cause of deafblindness at 4 to 17 per 100,000, is the clearest case, and those patients are routinely implanted [27].
Reported prevalence ranges across the literature. Each bar is the range as stated, not an interpolated estimate.
Prevalence ranges as reported in a study of refractive error, amblyopia, strabismus and low vision among deaf and hearing-impaired students, which also states the shared embryonic origin of retina and cochlea [26]. The underlying studies weight toward children and adolescents; we have not found an equivalent figure measured specifically in adult implant recipients, and we are not claiming one.
Combined hearing and vision loss climbs steeply in later life: about 1.5% of people aged 65 to 74, 2.6% at 75 to 84, and 10.8% past 85 [28]. The same age band where hearing loss concentrates is the band where the second sense starts going too.
A device that sits on the face can carry optical correction and audio on one mount. We are not building display features for their own sake, but the person we are designing for is more likely than not to need a lens in front of their eye regardless of what we do.
Only 16% of US adults aged 20 to 69 who could benefit from a hearing aid have ever used one [19]. That number, not the engineering, is what has killed assistive hearing hardware for forty years. People do not decline these devices because they do not work. They decline them because of what wearing one means.
Everything below runs on the glasses themselves. Nothing you say or hear is sent to a server. Six separate models share one chip on the frame, and the whole path from sound in your ear to text in your eye is built to stay under 120 ms. The badge on each one says how finished it is, not what it does in a clinic.
A causal, streaming separation network pulls the talker you are attending to out of overlapping speech and diffuse babble, on-frame, with lookahead capped at 32 ms. It descends from the line of work that took source separation from an offline trick to a real-time one [15][16].
Enroll the voices that matter, a partner, a child, a manager, in a few seconds each. Leo learns their embeddings and prioritizes them across rooms and days, and tags every caption with who is speaking so a crowded table reads as one attributable thread.
Gaze direction, head orientation and streaming diarization fuse into a single target estimate, so Leo locks onto the person you are looking at without you touching a thing. When confidence drops it widens rather than guesses, because a wrong lock is far more disorienting than a broad one.
A speech-translation model renders a talker in another language into the wearer's language, captioned in the display and optionally spoken. On-device and scoped to conversational domains, we are not claiming open-domain, unbounded translation, and we would distrust anyone who did.
Transcribed speech lands in the lower periphery of a monocular waveguide, each line tagged to the direction it arrived from and never placed over a talker's mouth. Text and sound are time aligned to fall inside the same audiovisual binding window, by design, not by luck.
Two cameras under the bridge read the wearer's signing space, an on-device pose model reads the hands, and the frame speaks. It gives a Deaf signer a voice with people who do not sign. It is the most valuable thing we could build and the least finished. Read the honest limits →
A modern cochlear implant restores access to sound with remarkable reliability. What it does not restore is the ability to pull one voice out of a room full of them. That failure is not a defect in the device, it is a consequence of what an electrode array can physically do inside a cochlea, and it has been stable in the literature for three decades.
people live with some degree of hearing loss; 430 million require rehabilitation today.
WHO, World Report on Hearing (2021) [1]
cochlear implant recipients worldwide, a population that has roughly doubled in a decade.
NIDCD implant statistics [2]
effective spectral channels, regardless of whether the array has 12 or 22 electrodes.
Friesen et al., JASA (2001) [3]
effective SNR gain available from visual speech cues alone at poor signal-to-noise ratios.
Sumby & Pollack, JASA (1954) [4]
People living with hearing loss worldwide, today versus the WHO's 2050 projection.
World Health Organization, Deafness and hearing loss fact sheet, drawing on the World Report on Hearing (2021). "Rehabilitation" counts moderate or higher loss in the better ear. [1]
Share of people with disabling hearing loss living in low- and middle-income countries.
WHO fact sheet [1]
Share of people older than 60 affected by disabling hearing loss.
WHO fact sheet [1]
What the hardware provides, against what listeners can actually use. Each bar is the range reported in the cited work, not an interpolated curve.
Asymptote of 8–10 channels: Berg et al., JASA 147(5), 2020, in 18 adult recipients on CNC words and AzBio sentences [11]. Continued gains to 20 channels for normal-hearing listeners on vocoded speech: Friesen et al., 2001 [3]. Electrode counts are the commercial range across current manufacturers.
An implant array sits in a conductive salt bath. Current injected at one electrode spreads to neighbouring neural populations, so physically adjacent contacts excite overlapping regions of the spiral ganglion. Adding electrodes does not add independent channels; it adds correlated ones.
Friesen and colleagues showed the consequence directly: normal-hearing listeners given noise-vocoded speech keep improving as channel count rises toward twenty, while implant users plateau around seven to ten and go no further [3]. Shannon's earlier work established the floor, intelligible speech needs surprisingly few channels in quiet [5]. Noise is where the gap opens.
Spectral resolution is exactly the resource you spend on separating concurrent talkers. With it degraded, implant users typically need a signal-to-noise ratio roughly 10–15 dB more favourable than a normal-hearing listener to reach comparable sentence recognition, and restaurants, classrooms and open offices sit well below that line.
Everything above this line comes from published work. This section does not. It comes from conversations with cochlear implant recipients about what actually goes wrong in a room, and it changed what we decided to build. We are stating it plainly as what it is: a small number of informal conversations, not a study, with no protocol and no sample size worth quoting.
They said they nod. They follow most of a conversation, lose a sentence, and rather than stop five people to ask, they guess and keep going. The cost is not a missed word, it is going home unsure whether the thing they agreed to was a question, a joke or a plan.
That failure is invisible from the outside and invisible on a test. A recipient scoring well on sentence recognition in a clinic can still be doing this at every dinner.
The other thing we heard repeatedly is that following a room is possible and expensive. The implant works. The person can do it. It costs so much concentration that they are wiped out afterwards, and the rational response is to go to fewer rooms.
An implant that succeeds on every clinical measure can still lose someone their social life through pure effort cost. That is a device outcome, and nothing in the standard battery captures it.
Both failures share a shape. Neither shows up in word recognition scores, which is what the field optimises, and both are caused by the same thing: the listener is doing separation work that the device should be doing. Effort is the product spec, not intelligibility.
It is also why we are not building captioning glasses. Reading a transcript solves comprehension and makes effort worse, because now you are decoding text and watching a face at the same time. If we ever measure a Leo that raises sentence scores while raising listening effort, we will report that as a failure, and we are stating that here so it is on record before we have data.
This is the finding our entire product thesis rests on, and it is not ours, it is well replicated. The visual channel in an implant user is not a fallback. It is load-bearing.
Percentage-point gain in word intelligibility from watching the speaker's face, at two signal-to-noise ratios. The size of the closed response set changes how much vision can contribute.
Sumby & Pollack, JASA 26(2), 1954 [4]. Spondee words, 129 listeners, each point pooled from 450 determinations. The paper reports the −30 dB gap as ranging "from 40 percent for the 256-word vocabulary to 80 percent for the 8-word vocabulary," and states that under noise-free conditions there is little difference between the two presentation conditions, plotted here as approximately zero.
Sumby and Pollack's 1954 measurement still sets the ceiling: at low SNRs, seeing a talker's face is worth as much as adding roughly 15 dB of clean signal [4].
Rouger et al. found implant recipients outperform normal-hearing controls at lipreading and fuse audiovisual speech more effectively, a durable reorganization, not a temporary strategy [6].
Audiovisual speech fuses inside a temporal binding window of roughly 200 ms, and asymmetrically, listeners tolerate audio lagging vision far better than the reverse [7].
AI glasses built for cochlear implant recipients. Leo does two things a sound processor cannot do on its own: it hears the room from a wider baseline than a single behind-the-ear microphone, and it renders what it hears into the wearer's visual field, where the evidence says it will actually be used.
A behind-the-ear processor gets one microphone and a head's worth of acoustic shadow. Leo distributes six MEMS capsules across the temples and brow bar, giving roughly a 140 mm horizontal baseline, enough aperture for meaningful directivity in the 500 Hz–4 kHz band where speech information concentrates.
A minimum-variance distortionless-response beamformer steers toward the attended talker; a second, wider-aperture estimate maintains a running map of competing sources so the system can switch targets without re-converging from scratch.
Amplification raises target and masker together. Directional gain does not, it improves the ratio. For a listener whose spectral resolution is capped by the physics of current spread, ratio is the only lever that still moves the needle.
Separation, diarization and transcription run on an on-frame NPU. No audio leaves the device. This is a privacy commitment, but it is a latency commitment first, a round trip to a phone, let alone a server, spends the entire budget before any processing begins.
The stack is a causal, streaming design end to end: no model in the chain is permitted to look ahead more than 32 ms, which rules out most published offline separation architectures and drives essentially every engineering trade-off downstream.
Our end-to-end design target from acoustic wavefront to implant stimulation is under 120 ms, chosen to sit inside the audiovisual binding window with margin [7]. Broadcast standards put the threshold of noticeable audio lag near 125 ms [8]; we treat that as a hard ceiling, not a goal.
Every stage in the pipeline below carries an allocated slice of that budget, and any model that cannot meet its allocation does not ship, regardless of accuracy.
Leo does not replace a sound processor and does not touch the implant's stimulation strategy. It presents itself as an external audio source over the accessory pathways implant manufacturers already expose, the same class of interface used by remote microphones and streaming accessories.
The processed stream is delivered as clean, pre-separated audio. The recipient's existing map, fitted by their audiologist, does the rest. Nothing about a Leo user's clinical relationship changes.
Three manufacturers account for the overwhelming majority of implants in the field, each with its own accessory protocol and its own telecoil and streaming behaviour. Leo abstracts these behind a single internal coupling layer so the acoustic stack never encodes vendor specifics.
A monocular waveguide renders transcribed speech near the lower periphery. Critically, each line is tagged to a spatial source, the system knows which direction a phrase arrived from, so the wearer knows who said it without breaking gaze to check.
Text is never placed over a talker's mouth. Given how much of the signal implant users extract from articulation [6], occluding the face to display a caption would cost more than the caption returns.
McGurk and MacDonald demonstrated that mismatched audio and visual speech does not degrade gracefully, it produces a confidently perceived third syllable that was never present in either stream [9].
A device that puts audio and text in front of the same person at slightly wrong times is not neutral; it manufactures errors. This is the single strongest argument for the latency budget being treated as a safety requirement rather than a performance one.
Everything else in Leo carries the room to the wearer. This carries the wearer to the room. Two downward-tilted global-shutter cameras under the bridge watch the signing space; an on-device pose and hand model reads it; the result is spoken aloud.
For a deaf signer talking to someone who does not sign, that removes the phone, the notepad and the interpreter from the loop. It is the most valuable thing we could build and the least finished, see the full write-up for what is actually hard about it.
American Sign Language is a language, with its own grammar, morphology and word order. It is not English on the hands. A system that maps signs to English words in sequence produces something between broken English and nonsense.
Six stages, one budget. Allocations below are DK-1 design targets, the budget we are writing software against. They have not been measured on hardware, because the hardware does not exist yet.
Six MEMS capsules sample at 48 kHz with sub-sample-synchronised clocks. Phase coherence across the array is the whole game; a drifting clock destroys directivity faster than any amount of self-noise.
Steered-response power localisation builds a running map of active sources around the wearer, updated every 16 ms. Head motion is compensated from the onboard IMU so the map lives in room coordinates, not head coordinates.
A causal, streaming separation network isolates the attended talker from concurrent speech and diffuse background, the descendant of a line of work that took source separation from an offline trick to a real-time one [15][16]. Lookahead is capped at 32 ms, which eliminates most published offline architectures outright.
Gaze direction, head orientation and streaming diarization, who is speaking, when, combine into a target estimate and the speaker tag each caption carries [17]. When confidence is low the system widens rather than guesses; a wrong lock is far more disorienting than a broad one.
Clean audio is streamed to the recipient's processor over the manufacturer's accessory pathway. Leo hands over a better signal; the existing clinical map converts it to stimulation exactly as it always has.
Transcription and speaker attribution reach the waveguide on a deliberately separate path, time-aligned to the audio stream so text and sound land inside the same binding window.
Four milliseconds of headroom is not comfortable, and we say so plainly. Reclaiming it is the central engineering objective of the DK-2 revision.
A cochlear implant solves one direction of the conversation. It does nothing for the other. A deaf signer speaking to someone who does not sign still needs a phone, a notepad or a human interpreter. Leo's outward channel is our attempt at removing that, cameras watch the signing space, an on-device model reads it, and the frame speaks. In a group, that is the difference between answering in real time and waiting for a turn that never comes.
Two global-shutter cameras sit under the bridge, tilted down toward the signing space in front of the torso. Global shutter matters here: signing is fast, and a rolling shutter skews a moving hand badly enough to change the handshape.
A hand and upper-body pose model runs on the same NPU as the audio stack, producing joint positions rather than raw video. Frames are discarded immediately, nothing that could reconstruct a conversation is retained or transmitted.
Fingerspelling and a constrained phrase set, decoded as a sequence. This is deliberately narrow. Continuous, open-domain sign translation is not a solved problem and we are not going to pretend otherwise.
Synthesised through a small frame speaker, or to a paired phone when the wearer would rather the voice came from the table than from their head. The wearer picks the voice, and the wearer confirms before anything is spoken.
Non-manual markers, brow position, head tilt, mouth morphemes, carry grammatical meaning in ASL: negation, topic marking, and the difference between a yes/no and a wh- question. A frame worn by the signer cannot see the signer's own face, so that information has to be inferred from the hands alone. This is the hardest structural problem in a first-person mount, and it is why we scoped to fingerspelling first.
From a head-mounted camera looking down, hands cross, overlap and foreshorten continuously. A third-person camera sees a signing space that a first-person camera sees edge-on. Most published sign-recognition results come from front-facing cameras at conversational distance, which is a materially easier problem than the one we have chosen.
Speech recognition had decades of transcribed audio before it worked. Sign language has orders of magnitude less data, and the largest continuous corpora are small next to what speech models train on. Progress is real and fast, but the starting point is not comparable.
| Dataset | Scope | Scale | Task |
|---|---|---|---|
| WWLASL | isolated | 2,000 signs | Word-level ASL recognition, over 21,000 clips. The easier problem, and the one closest to fingerspelling. |
| HHow2Sign | continuous | 11 signers | Continuous ASL over a vocabulary of more than 16,000 English words. Much harder, much closer to real conversation. |
| PPHOENIX-2014T | continuous | one domain | German Sign Language weather broadcasts, the field's standard translation benchmark, and narrow by construction. |
WLASL: Li et al., WACV 2020 [12]. How2Sign: Duarte et al., CVPR 2021 [13]. RWTH-PHOENIX-Weather 2014T, introduced for neural sign language translation by Camgöz et al., CVPR 2018 [14]. All three are third-person recordings; none is first-person, which is the gap our own data collection has to close.
The figures below are drawn from published work on cochlear implant performance, not from Leo. We publish the baseline we are trying to move before we publish anything about moving it.
| Listening condition | Channels available | Channels used | Ceiling behaviour | |
|---|---|---|---|---|
| 1 | NNormal hearing, unprocessed | , | full | No ceiling; performance limited by the masker, not the listener. |
| 2 | VNormal hearing, 20-ch vocoder | 20 | ~20 | Continues improving with channel count up to twenty. |
| 3 | VNormal hearing, 8-ch vocoder | 8 | 8 | Approximates implant performance in quiet; diverges in noise. |
| 4 | CIImplant, 22-electrode array | 22 | ~8 | Plateaus near eight regardless of electrodes active. |
| 5 | CIImplant, 12-electrode array | 12 | ~7 | Statistically indistinguishable from the 22-electrode case. |
| Environment | Typical SNR | CI shortfall | Intelligibility for implant users | |
|---|---|---|---|---|
| 1 | QQuiet room, one talker | +20 dB | none | |
| 2 | HHome, television on | +10 dB | ~5 dB | |
| 3 | OOpen-plan office | +5 dB | ~10 dB | |
| 4 | CClassroom, group work | 0 dB | ~12 dB | |
| 5 | RRestaurant, multi-talker babble | −5 dB | ~15 dB |
Channel counts follow Friesen et al. [3] and Shannon et al. [5]. Environmental SNRs are representative ranges for the listed settings, and the intelligibility bars illustrate the direction and magnitude reported across the speech-in-noise literature rather than any single pooled dataset. Individual outcomes vary enormously, duration of deafness, age at implantation and residual hearing all dominate.
Developer kit configuration. Every figure below is a design target for a frame we have not built. Nothing here has been measured, certified or clinically verified. Treat this table as the specification we are designing to, which is the only thing it is.
Hearing is one of the biggest unaddressed health burdens in the world and one of the smallest device markets relative to it. The gap between those two facts is the whole opportunity, and it is also the reason to be careful: a market that has looked obviously large for thirty years and stayed small is telling you something about adoption, not about need.
annual global cost of unaddressed hearing loss, across health care, lost productivity and societal cost.
US adults who could benefit from a hearing aid. About 16% of those aged 20 to 69 have ever used one.
NIDCD, Quick Statistics About Hearing [19]
cochlear implants registered worldwide as of December 2019; 118,100 in US adults, 65,000 in US children.
FDA figures via NIDCD [2]
year-on-year growth in smart glasses shipments through 2025, with one vendor holding over 70% of the category.
Counterpoint Research [21]
Published estimates for the same market in the same year, plotted as the span between the lowest and the highest. Each span is roughly a factor of two.
Cochlear implant estimates for 2025 collected from published report summaries by Grand View Research, IMARC, Vantage Market Research, Fortune Business Insights, Research Nester and Mordor Intelligence; hearing aid estimates from eight comparable reports. These are commercial research products, not peer-reviewed work, and their methodologies are not public. Definitions differ on whether over-the-counter devices, hearables and service revenue are counted. The spread is the finding [25].
US utilisation among people meeting traditional audiometric candidacy criteria. Under expanded criteria it falls to about 2%.
Nassiri, Sorkin & Carlson, Otology & Neurotology (2022) [20]
US adults aged 20 to 69 who have ever used a hearing aid, among those who could benefit. It reaches 30% only over age 70.
NIDCD, Quick Statistics About Hearing [19]
Because the category boundary is unsettled. A hearing aid was a regulated medical device until the US over-the-counter rule took effect in 2022, and since then the line between a hearing aid, a hearable and a pair of earbuds with a hearing feature has been a matter of opinion. Firms that count earbuds report figures around $15B. Firms that count only fitted devices report around $9B.
We are quoting the disagreement rather than picking the flattering end of it. A number you cannot reproduce is not evidence, and the honest version of this slide is that the addressable device market is somewhere in the low tens of billions and growing at high single digits.
Leo's first customer is an adult cochlear implant recipient in a high-income market. Cumulative US adult implants are 118,100, and cumulative implants are not living active users, so the real number is smaller. Call the reachable beachhead on the order of 100,000 people in the US and perhaps three times that across comparable markets.
Assistive glasses have failed on stigma before. Smart glasses shipments roughly doubled year on year through 2025, and someone else's marketing budget is normalising a camera and a microphone array on a face. A device that looks like what everyone else is already wearing does not carry the same cost.
Real-time speaker separation good enough for a live conversation is recent, and the models that made it practical are small enough to run causally on a frame rather than in a datacentre [15][16]. That is a capability change, not a packaging change.
Every major implant manufacturer already ships an accessory streaming pathway. We do not have to displace an incumbent or win a clinical channel to reach a recipient. We have to be compatible with the one they already have.
Glasses for hearing loss are not a new idea and we are not the first people here. It is worth being precise about what already exists, because the thing that separates Leo from all of it is narrow and easy to state: every product below either makes the whole room louder or turns it into text. None of them decide which voice you want and send only that one into a cochlear implant.
| What exists | What it does | Who it is for | Why it does not solve this |
|---|---|---|---|
| Nuance Audio EssilorLuxottica |
Open-ear directional amplification built into a normal-looking frame. FDA cleared as an over-the-counter hearing aid and on sale in the US. | Perceived mild to moderate hearing loss. | Amplification, not separation. Making a room with four talkers louder gives an implant recipient four louder talkers. It is also not a route into an implant. |
| AirPods Pro hearing aid feature Apple |
FDA-authorised hearing aid function in a mass-market earbud, with a self-administered hearing test. | Mild to moderate loss, in-ear. | Same amplification limit, plus it occupies the ear. An implant recipient's hearing does not run through their ear canal. |
| Captioning glasses Xander, XRAI and others |
Live speech-to-text projected on the lens. Xander runs on the device without a phone; XRAI runs off a connected phone. | Severe loss and Deaf users who read English fluently. | Text instead of sound, and it costs you the talker's face. Reading captions pulls your eyes off the mouth, and the visual speech cues you lose are worth up to 15 dB of effective SNR [4]. This is the tradeoff our whole thesis is built on. |
| Manufacturer remote mics Cochlear, Advanced Bionics, MED-EL |
A clip-on microphone that streams one talker directly into the implant over the accessory pathway. | Cochlear implant recipients. Widely prescribed. | This is the real incumbent and the one to beat. It works, but somebody has to physically wear it, so it solves one known talker in advance and does nothing for a table of five. |
The remote microphone is the honest baseline, not Nuance Audio and not captions. A recipient at a dinner table today has one option that genuinely works: hand a microphone to one person and give up on everyone else. Leo's claim is that an array on your own face can do the choosing, with no object to hand over and no conversation to interrupt.
That is a narrow claim, and it is falsifiable. If our separation does not beat a clipped-on remote mic in a five-talker room, we do not have a product, and that comparison is the first measurement we intend to publish.
We are not competing with Nuance Audio or Apple. They serve mild to moderate loss, a population perhaps a hundred times larger than ours, and they serve it well. If anything, they help us: every pair of hearing glasses sold makes the form factor more ordinary and lowers the stigma cost of the one we are building.
Two people so far, paired on purpose: one who has spent years turning a condition he lives with into working assistive technology, and one who brings the clinical depth that hearing work demands. Between us we have the two things this problem is hard to attempt without: direct access to recipients and clinicians, and a founder who has already shipped assistive technology for a condition nobody was going to solve for him.
Hardware & systems
Kiro studies computer science, economics and data science at MIT. He was born with congenital nystagmus, a condition with no established cure in that form, and instead of waiting for one he built computer-vision software that counter-oscillates on-screen text to match the motion of the eye, holding reading still for people who live with it. It began with his own sight and his sister's data, and earned two Silver awards at a regional science and engineering fair.
The Ear Company comes from the same instinct: not assistive technology in the abstract, only the version that actually reaches the person who needs it. kiro.city ↗
Clinical & language
Finn is an incoming computer science undergraduate and works at Stanford on clinical speech-language pathology and NLP. His regular work with deaf children brings exactly the clinical depth a hearing product demands, the part engineering alone cannot supply.
Kiro and Finn have known each other for two years. The split is deliberate: hardware and systems on one side, clinical and language science on the other, with the people Leo is for kept at the centre of both.
Practically, this is why the conversations behind the section above happened at all, and it is the route to the academic audiology partner the 2027 studies depend on. For a company whose hardest gate is recipient access rather than engineering, that is the part of the team that is hard to hire for.
We are a hardware company in a clinical adjacency, which means the honest version of our timeline is longer than the exciting version. This is the honest one.
Where we actually are. The separation and sign models run on a laptop against recorded multi-talker audio. The DK-1 array geometry is designed and not yet built, so nothing has been characterised in a chamber and no latency figure on this site has been measured on hardware.
The whole device, assembled and running unaided: six-microphone array, separation and sign both on-frame, own battery, coupled into an implant. This is the milestone that turns every design target on this page into a measurement, and it is what the raise is for. Until it lands, the honest description of Leo is software plus a drawing.
Bench interoperability across the three major implant accessory pathways, with a published compatibility matrix. No performance claims until this is stable.
Structured speech-in-noise testing with adult implant recipients under IRB oversight, run with an academic audiology partner. Protocol pre-registered; results published whichever way they land.
Longitudinal at-home use. Bench SNR gain and real-world benefit are different quantities, and the gap between them is where most assistive hearing technology has historically died.
Published sources for every empirical figure cited above. Where we state a design target rather than a finding, it is labelled as such in the text.
We build AI hardware for bionic enhancement. We started with hearing because it is the one sense where a widely deployed neural implant already exists, has been refined over four decades, and is still bottlenecked by something a well-designed external device can genuinely help with.
We build things that attach to the clinical world rather than replacing it. No recipient should have to abandon a working map, a trusted audiologist, or a device they have spent years learning.
Conversation is the most sensitive data a person produces. Leo processes it on the frame and keeps it there. That constraint shapes the architecture rather than being bolted onto it.
We put the literature we are trying to beat on our own website before we put up a single number of our own. If our results do not clear that bar, that will be visible here too.
The Ear Company, LLC is a limited liability company headquartered in San Francisco, California. Hardware, acoustics and machine learning are in one room, which for a latency-bound product is less a culture choice than an engineering requirement.
Join the waitlist for early access to the DK-1 developer kit and the occasional, honest update on where the engineering actually stands. No hype, no spam, unsubscribe in one click.
Audiologist, researcher or press? Email us directly ↗
If you invest: we are raising a pre-seed to build a fully working Leo, the whole device on a face rather than a bench, and then measure it in a real room against a clipped-on remote microphone, the incumbent we actually have to beat. That comparison is the next thing that will appear on this page, whichever way it lands. We are happy to walk through the off-frame demo and the latency budget on a call before you decide anything.
