Nine and a Quarter Milliseconds
Meta's FDA filing puts its glasses hearing loop at 9.25 ms. In an open ear that delay is audible, and its costs belong to the frame, not firmware.
Sit across a loud table from someone while wearing amplifying glasses and their voice reaches your eardrum twice: straight through the air into an ear canal nothing blocks, then a few milliseconds later from a speaker on the temple. The ear adds the two. Copies of a sound offset in time cancel at frequencies spaced one over the delay apart, so a millisecond of lag puts a notch every kilohertz and nine milliseconds puts one about every hundred hertz, straight through speech. An earbud's tip turns the direct copy down. Glasses leave the canal open by design, so the direct copy arrives at full strength.
Part 3 of six on the Connect 2026 documents; part 2 followed the puck's memory bus to Meta's cloud.
Meta's FDA filing for AI-glasses hearing enhancement, 510(k) K254044, lists a 9.25-millisecond latency against a 3.15-millisecond median for its Apple AirPods predicate. In an open ear, Meta's 9.25-millisecond delay is past the 5 to 6 milliseconds Stone and Moore's 2008 open-canal study found just acceptable. By my count of FDA-authorized hearing-aid software, Luxottica's and Meta's open-ear glasses file latencies near 9 milliseconds, Apple's and Samsung's in-canal earbuds near 3. Meta designed its speech-in-noise hearing enhancement for at least 3 hours of continuous use. Nuance Audio, glasses built around hearing, also sits near 9 milliseconds, so getting open-ear glasses under 5 milliseconds reads as a frame problem, not a firmware update.
Which part of the glasses did the FDA clear?
Only the software, cleared July 13 as an over-the-counter, self-fitting hearing aid for adults with perceived mild to moderate loss. Meta's 510(k) summary calls it Hearing assist: an iPhone app plus firmware for glasses it never names.
In the clearance trial, 142 adults at three US sites wore the glasses at home for 14 days. Self-fitters gained 4.8 dB over their unaided scores on a speech-in-noise test, though with a 5.3 dB standard deviation a normal spread leaves about one in five with no benefit (my arithmetic; the filing reports no distribution).
The filing's Table 1 has its own catch. Apple's row reads "Median Latency: 3.15 msec"; Meta's reads "Device Latency: 9.25 msec", with no method or spread. The over-the-counter rule, 21 CFR 800.30, caps latency at 15 ms and asks only for a method accurate and repeatable to within 1.5 ms. Meta uses 62 percent of that cap (9.25 / 15), and even if its measurement read the full 1.5 ms high, 7.75 ms is still past the open-canal 5 to 6 ms. FDA checked the cap, not how 9.25 ms sounds.
Why does 9.25 ms matter when the ear is open?
Because the ear hears both copies, and at 9.25 ms the pair is already past the delay open-canal listeners rate just acceptable. The delay also sets how the pair sounds: cancellations every 1 / 0.00925 s = 108 Hz, about 70 notches across Meta's cleared 232 to 7,809 Hz band, against 317 Hz spacing at Apple's 3.15 ms (computed from the filed numbers, not measured). Meta's Ravish Mehra told TechRadar the loop has to be fast "because otherwise you start hearing echoes or start hearing distortions". At 343 m/s, 9.25 ms is 3.17 m of travel, the delay of a reflection off a wall about 1.6 m away. That is too short to hear as an echo, so the ear fuses the copies into one colored sound: the distortion, not the echo.
Notch depth depends on the level gap. Equal copies cancel completely; 10 dB apart, the ripple is about +2.4 / -3.3 dB, and 20 dB apart it stays within a decibel (computed). Small gain, at the mild end of Meta's indication, keeps the copies near equal.
Stone, Moore and colleagues' 2008 open-canal study found delays of "about 5 and 6 msec" just acceptable for simulated losses starting at 2 and 1 kHz. Its data blunts my comb argument: filtering out the ripples helped little, and with compression on, disturbance rose with gain. So I blame the delay; the comb is how you recognize it.
The best open-ear witness predates the clearance: HearingTracker, reviewing Gen 2's uncleared Conversation Focus, warned of a "slightly metallic" or phasey quality from "the mix of open-ear and amplified audio".
Two glasses near 9, two earbuds near 3
By my count, five products hold authorization under SCR, the FDA's product code for air-conduction hearing-aid software (four 510(k) clearances and Apple's De Novo grant), and four file a latency. Luxottica's Nuance Audio clearance, the glasses precedent Meta's filing cites, lists 8.8 ms; Apple's AirPods feature, 3.15; Samsung's Galaxy Buds feature, cleared July 31, 2.7. The glasses' filed bands also top out lower, at 6.3 and 7.8 kHz against the earbuds' 9 and 10. The order is backwards from what the ear needs: the earbuds, whose tips already turn the direct copy down, file the short delays, and the glasses, where it arrives at full strength, file the long ones.
Mehra named both costs to TechRadar: sound from the temple speakers will "feed back into the microphones" across open air, and the processing has to fit "a very small envelope" of onboard compute. The filing states no sample rate, but Meta's 7,809 Hz edge sits 191 Hz under 8 kHz, the ceiling for audio sampled at 16 kHz. At 16 kHz, 9.25 ms is 148 samples for capture, beamforming, gain, feedback control and playback (inferred, not filed).
On my reading, firmware can buy back milliseconds only by dropping stages that earn the benefit or keep the loop stable; shorter blocks need more compute than that envelope holds. Nuance sits within half a millisecond of Meta.
Three hours, on the frame
The filing's battery line:
Hearing assist is designed to last at least 3 hours of continuous use on fully charged Smart Glasses.
Mehra told CNET the mode lasts about 4 hours on Meta's recent glasses and longer on the Audio pair. Meta rates Gen 3 at up to 9 hours and Audio at up to 12, so the filed floor is a third of one and a quarter of the other. At Mehra's 4 hours, hearing mode draws about 2.25 times the average power of whatever use Meta rates for Gen 3 (9 / 4). At that rate a two-hour dinner spends half the charge the always-on assistant was counting on.
A listening assistant can buffer audio and sleep between bursts; a 9.25 ms loop keeps microphones, speaker and processor running all session. The filing puts that loop in the glasses' firmware; after the display link in part 1, it is the second thing physics pins to the face.
With latency a wash, Nuance loses buyers who need help at dinner rather than all day. Against Meta's $150 unlock, Nuance's best argument is runtime: about 6 hours of hearing use in its clearance, and about 10, by EssilorLuxottica's own claim, for the $829 Nuance Audio Plus, sold through optical and hearing-care professionals since September 29. That defense lasts exactly as long as Meta's three hours do.
This reading fails if, before January 1, 2028, an FDA filing, a Meta spec page or an independent lab measurement puts hearing enhancement under 5 ms on any Meta glasses already on sale on October 31, 2026. It also fails if Meta or an independent test shows it running continuously for at least half the frame's rated battery, 4.5 hours on Gen 3 or 6 on Audio; Mehra has already told CNET that Audio lasts longer than 4 hours, so that is the likelier way to lose.