From Hearing Aid to Hunter-Killer:

 

The Edge-AI Acoustics Convergence

in the style of U.S. Naval Institute Proceedings — strategy and systems analysis, August 2026

---

Bottom Line Up Front

A $250 earbud and a $3-million unmanned submarine now share a core engineering problem: how to separate a wanted acoustic signal from noise, classify it, and act on it — in real time, on a starvation power budget, with no human in the loop and no link to the cloud. The commercial hearing-aid and consumer-audio industries have quietly won the race to run neural networks continuously at milliwatts. The Navy needs precisely that capability afloat and, more urgently, submerged, aboard the extra-large unmanned undersea vehicles (XLUUVs) it is now committing to buy in quantity. The technical convergence is real and the Navy's science-and-technology community sees it clearly; the funding threads run through the Office of Naval Research, Program Executive Office (PEO) Undersea Warfare Systems, the air anti-submarine warfare (ASW) program office, and DARPA. The limiting factor is not the algorithm or the silicon — both are maturing fast in the commercial base — but the Navy's acquisition system, which the Government Accountability Office (GAO) has repeatedly faulted, and the clock being set by a peer competitor making bold, if unverified, claims about AI-enabled ASW. The author contends that the Navy should treat edge-AI acoustic inference as a portable, platform-agnostic payload layer, harvest the commercial "AI-per-watt" silicon base rather than reinvent it, guard its acoustic training data as the true crown jewel, and fix the business case before it scales the hull.

---

An unlikely emblem

On 24 August 2026, the Swiss hearing-aid maker Phonak — a Sonova brand — launched a device called EON. It is an improbable place to begin a discussion of undersea warfare, and that is exactly the point. EON runs two neural-network systems at once: one, Spheric Speech Clarity 3.0, uses a deep neural network (DNN) to pull a talker's voice out of background noise at the waveform level; the other, AutoSense OS AI 8.0, classifies the acoustic scene and reconfigures the processing chain to match it. It does this continuously, all day, on a custom low-power processor the company says draws roughly 37 percent less power than its predecessor while shrinking the package 25 percent — a device that must disappear behind an ear and never run hot.

Strip away the consumer-health packaging and the engineering statement underneath is stark: learned acoustic source separation plus real-time scene classification, executed at milliwatts on purpose-built silicon, untethered from any datacenter. That sentence also describes the processing an autonomous undersea vehicle must perform to earn its keep in the ASW fight. The hearing aid is not a weapon system. It is an existence proof — evidence that the hardest part of the problem, sustained neural inference inside a brutal size-weight-and-power (SWaP) envelope, has been solved at commercial scale and consumer price.

The same problem — but not the same ocean

Naval professionals should resist the temptation to over-read the analogy, because the physics diverge sharply below the waterline. Air-conducted speech is a forgiving medium. The undersea channel is not. A 2026 systematic review in the Journal of Field Robotics catalogs the difference plainly: underwater acoustic processing must contend with multipath propagation, Doppler shifts, ambient noise, reverberant littoral zones, low-observable targets, and time-varying interference that together degrade the classical beamformers, matched filters, and deterministic classifiers on which legacy sonar was built. The frequencies are lower, the arrays larger, the propagation ranges longer and stranger, and — decisively — the training data are scarce, expensive, and largely classified, in contrast to the ocean of labeled speech that trained the hearing aid.

So the transfer is at the level of method and hardware, not domain. What crosses over is the toolkit: DNN-based separation replacing matched filters, learned classifiers that can build templates for never-before-heard targets, sensor fusion, and above all the SWaP-optimized inference silicon that makes any of it possible on a small, unattended platform. The peer-reviewed defense literature already reports hybrid AI sonar models achieving classification accuracies above 90 percent with improved signal-to-noise ratio and reduced false alarms — the direction of travel is not in doubt, only its operational maturity.

The dual-use flow has reversed

Here is the part that should reorient how the acquisition community thinks. For most of the Cold War, the technology flowed from defense to the commercial world. Adaptive beamforming, matched-field processing, and towed-array signal processing were Navy and national-laboratory achievements that trickled out to seismic surveying and, eventually, consumer audio. That current has now reversed. The frontier of low-power neural inference is being pushed hardest by the commercial edge-AI industry, because the economic prize — putting intelligence into every phone, camera, and earbud without a cloud connection — is enormous.

The evidence is on the trade-show floor, not the test range. Qualcomm now ships neural processing units in hundreds of millions of Snapdragon chips a year; Google fields Edge tensor-processing units for on-device inference; and a cohort of specialists — BrainChip's Akida, Syntiant, and others — build brain-inspired, event-driven parts that perform real-time inference within milliwatt power budgets. The canonical benchmark these companies chase is keyword spotting: recognizing a spoken word, on-chip, at negligible power — the direct commercial cousin of passive acoustic target recognition. Independent research finds that conventional DNNs on edge processors can consume one to three orders of magnitude more energy than neuromorphic approaches for equivalent throughput, which is why "maximum AI performance per watt" has become the industry's organizing principle. Apple's earlier absorption of the edge-AI startup Xnor.ai and Qualcomm's recent acquisitions of Edge Impulse and Arduino show the incumbents buying their way onto that frontier.

For a Navy trying to field undersea autonomy, this is a strategic gift: the SWaP problem that would once have consumed a decade of in-house development is being solved, at scale and under competitive pressure, by a commercial base the service can exploit. The task is integration and adaptation, not invention.

What the Navy is actually funding

The service is not asleep to this. The activity is spread across the S&T and acquisition enterprise, and it is accelerating.

On the signal-processing core, the Navy in October 2025 awarded Metron a contract to develop advanced sonar computing and sensor processing for ASW and undersea surveillance, explicitly emphasizing artificial intelligence, machine learning, and high-performance sonar for rapid threat detection — with a parallel award to Serco focused on AI, machine learning, and predictive analytics for signals analysis and decision support. At the fleet-demonstration level, Lockheed Martin and the Navy used the Rim of the Pacific (RIMPAC) 2026 exercise to demonstrate an AI/machine-learning capability that rapidly updates ASW acoustic classification models. Lockheed's SensorMAX, built on what the company calls a Spectral Foundation Model, is designed to push secure, distributed AI-model updates without reliance on continuous connectivity — a requirement written, in effect, for a disconnected undersea platform.

The requirements language is even more revealing in the small-business pipeline. A current Navy Small Business Innovation Research topic from the air ASW systems program office (PMA-264) seeks technologies to reliably detect, classify, track, and localize submarines and unmanned undersea vehicles via passive sensors, hosted on the acoustic processor of a manned or unmanned aircraft — and it points specifically to advanced methods for generating matched filters or templates for never-before-seen targets. That is the generalization problem at the heart of modern machine learning, stated in ASW terms. PEO Undersea Warfare Systems, for its part, funds a Submarine Combat System Improvement (Advanced) line that develops sonar, combat-system, and sensor-processing software in support of acoustic superiority and technology insertion.

The platform-side push is just as clear. DARPA's Manta Ray program set out to demonstrate, among other things, novel energy management and harvesting for long-endurance undersea operation and new low-power means of underwater detection and classification of hazards and counter-detection threats — completing full-scale in-water testing in 2024. DARPA's follow-on "Deep Thoughts" solicitation, released in April 2026, continues the work on autonomous-underwater-vehicle designs, embedded subsystems, and mission engineering. And in the budget, the Navy's Combat Autonomous Maritime Platform (CAMP) effort carried a roughly $98-million fiscal-2027 request to accelerate the Orca XLUUV line.

The platform that needs it most

Why does this matter more for undersea autonomy than for any surface combatant? Because a submerged, unmanned vehicle is the one platform that cannot phone home. The Navy's Boeing-built Orca XLUUV is designed for months-long missions and ranges reported up to 6,500 nautical miles; in July 2026 it completed the program's first 1,000-nautical-mile Pacific transit. A vehicle of that endurance, operating in a communications-denied environment, cannot stream raw hydrophone data to a ship for a sailor to interpret. It must detect, classify, and decide onboard, or it is merely an expensive drifting sensor recording for a post-mission data dump. A Navy unmanned-systems director, Captain Matt Lewis, has put the core difficulty plainly: once a vehicle submerges, it must manage the air-water interface and command-and-control latency without the human judgment a crewed submarine brings to the fight.

That is the operational reason the hearing-aid comparison is more than a rhetorical flourish. The milliwatt, always-on, untethered inference that consumer acoustics has perfected is the exact enabling capability an XLUUV needs to convert endurance into effect. Edge-AI acoustic processing is what turns a long-range hull into a hunter.

The oversight paper trail — the real limiting factor

If the algorithm and the silicon are the good news, the acquisition record is the caution. The paper trail here is not courtroom litigation but the far more consequential oversight record of the GAO and the Congressional Research Service — and it is unflattering. GAO's September 2022 assessment (GAO-22-105974) found the Orca XLUUV at least three years late and 64 percent — roughly $242 million — over its original cost estimate, and faulted the Navy for pursuing an "emergent need" without a sound business case. By GAO's June 2025 weapon-systems assessment, the Navy had spent on the order of $885 million and it was, in the auditors' words, "unclear" whether the XLUUV would even transition to a program of record, because there were no clear requirements the vehicle could meet within budget constraints.

The Navy pressed ahead anyway. Its May 2026 shipbuilding plan moved Orca from prototype to program of record, funding two vehicles in fiscal 2027 and sixteen across the future-years defense program, with $135.8 million requested in fiscal 2027. Reasonable officers can disagree about whether that is boldness or sunk-cost momentum. But the lesson for the edge-AI acoustics enterprise is unambiguous: the binding constraint on fielding this capability is not whether a DNN can classify a contact at milliwatts — commercial industry has answered that — but whether the Navy can write disciplined requirements, structure a defensible business case, and avoid welding immature autonomy to an over-budget hull under schedule pressure. Capability aspiration is not fielded capability, and the auditors have been saying so for four years.

The clock

The urgency is external. In September 2025, Chinese researchers led by a senior engineer at the China Helicopter Research and Development Institute published, in a peer-reviewed journal, an AI-driven ASW concept they claim can fuse sonar, radar, magnetic-anomaly, and oceanographic data into a real-time picture and drive tactical recommendations — with an asserted success rate around 95 percent. Those figures are laboratory claims, not demonstrated fleet performance, and should be read with professional skepticism. But they do not stand alone. Reuters reported in March 2026 on a multi-year Chinese ocean-floor mapping effort across the Pacific, Indian, and Arctic oceans that would directly feed acoustic-propagation prediction for both submarine concealment and ASW. Chinese firms have displayed submarine-launched autonomous vehicles sized for 260-mm and 533-mm torpedo tubes, and open-source reporting in late 2025 described PLA Navy AI-enabled undersea drones said to operate below 90 decibels, execute zero-radius turns, datalink with one another, recharge at submerged stations, and run without an operator link.

Whether or not any single claim survives scrutiny, the pattern is the strategically relevant fact: the peer competitor is racing on precisely the axis this article describes — AI-enabled undersea acoustic sensing and untethered autonomy. The convergence is not a curiosity the United States can study at leisure. It is a contest.

Recommendations

Four propositions follow for the naval professional.

Treat inference as a portable payload, not a platform feature. The acoustic-AI processing layer should be abstracted from any one hull — Orca today, a glider or a torpedo-tube vehicle tomorrow — so that a classifier proven on one platform migrates to the next by design. Lockheed's disconnected-update model at RIMPAC 2026 points the way.

Harvest the commercial silicon base. The Navy should aggressively adapt commercial edge-AI and neuromorphic inference parts rather than fund bespoke undersea processors from scratch. The dual-use current now runs commercial-to-defense; the service that exploits it fastest wins the SWaP argument.

Guard the data, not just the device. In a domain where labeled acoustic signatures are scarce and classified, the training data — not the chip — is the true crown jewel and the genuine barrier to entry. Curating, protecting, and expanding the Navy's undersea acoustic datasets is a first-order investment, not an afterthought.

Fix the business case before scaling the hull. GAO's four-year record is a warning, not background noise. The most sophisticated onboard classifier in the world cannot rescue a program with no affordable mission it is sure it can perform.

The hearing aid behind a grandparent's ear and the hunter-killer prowling a contested strait are, at the level of silicon and mathematics, solving the same problem. The Navy did not create that convergence, and it will not own it. The question is whether the service can move at the speed of the commercial frontier it now depends on — and get its own acquisition house in order before an adversary converts the same physics into an advantage.

---

Sources

Commercial edge-acoustic AI (the "hearing aid" baseline)

1. Sonova International. "New EON hearing aid portfolio strengthens Phonak's leadership…" 24 Aug 2026. https://www.sonova.com/new-eon-hearing-aid-portfolio-strengthens-phonaks-leadership-in-speech-understanding-scene-intelligence-and-made-for-all-connectivity/
2. HearingTracker (Karl Strom). "Phonak Eon Debuts Smaller Sphere, Auracast, and WindBlock." 24 Aug 2026. https://www.hearingtracker.com/news/phonak-eon-debuts-smaller-sphere-auracast-and-windblock

Edge-AI / neuromorphic silicon convergence

3. Edge AI and Vision Alliance. "AI at the Edge: Low Power, High Stakes." 20 Nov 2025. https://www.edge-ai-vision.com/2025/11/ai-at-the-edge-low-power-high-stakes/
4. Woodside Capital Partners. "AI at the Edge: Low Power, High Stakes." 3 Nov 2025. https://woodsidecap.com/ai-at-the-edge-low-power-high-stakes/
5. Brightfield, S. (BrainChip). "How Neuromorphic Chips are Revolutionizing the Edge." All About Circuits, 30 Apr 2026. https://www.allaboutcircuits.com/industry-articles/how-neuromorphic-chips-are-revolutionizing-the-edge/
6. "Energy-Efficient Neuromorphic Computing for Edge AI…" arXiv, Feb 2026. https://arxiv.org/html/2602.02439v1
7. Mordor Intelligence. "Neuromorphic Chip Companies — Key Players." Jan 2026. https://www.mordorintelligence.com/industry-reports/neuromorphic-chip-market/companies

Underwater acoustic AI — technical literature

8. Das, S., & Pandey, A. "Challenges and Advances in Underwater Sonar Systems and AI-Driven Signal Processing for Modern Naval Operations: A Systematic Review." Journal of Field Robotics 43:899–931, 2026. https://onlinelibrary.wiley.com/doi/10.1002/rob.70077
9. Dipo Andimuharrom et al. "Enhancing ASW Capabilities Through Adaptive AI-Driven Sonar Signal Processing." Int'l Journal of Educational Technology Research 4(2):131–150, 2026. https://journalijetr.my.id/index.php/ijetr/article/view/6

U.S. Navy / DoD programs and demonstrations

10. Military & Aerospace Electronics. "Navy picks Metron for advanced research in sonar signal processing for ASW." 22 Oct 2025. https://www.militaryaerospace.com/computers/article/55324617/anti-submarine-warfare-asw-sonar-signal-processing
11. Military & Aerospace Electronics. "Navy chooses Serco for research in advanced sonar signal processing for ASW." https://www.militaryaerospace.com/computers/article/55274180/sonar-signal-processing-for-anti-submarine-warfare-asw
12. Military & Aerospace Electronics. "Lockheed demonstrates AI-enabled sonar classifier updates during RIMPAC." 10 Aug 2026. https://www.militaryaerospace.com/sensors/article/55396699/lockheed-demonstrates-ai-enabled-sonar-classifier-updates-during-rimpac
13. Navy SBIR (PMA-264). "Direct to Phase II: Passive Acoustics Sonobuoys Intelligence and Machine Learning" (Topic N251-D01). https://navysbir.us/n25_1/N251-D01.htm
14. DefenseScoop. "DARPA tests undersea Manta Ray drone prototype, looks to transition tech to Navy." 1 May 2024. https://defensescoop.com/2024/05/01/darpa-manta-ray-northrop-grumman-uuv-testing/
15. DefenseScoop. "DARPA shares 'Deep Thoughts' solicitation for autonomous underwater drones." 24 Apr 2026. https://defensescoop.com/2026/04/24/darpa-autonomous-underwater-vehicle-auv-program-deep-thoughts/

UUV platforms and budget

16. Naval News. "U.S. Navy Orca XLUUV Completes 1,000-Mile Pacific Transit." Aug 2026. https://www.navalnews.com/naval-news/2026/08/u-s-navy-orca-xluuv-completes-1000-mile-pacific-transit/
17. Boeing. "XLUUV (Orca)" platform page. https://www.boeing.com/defense/autonomous-and-unmanned-systems/xluuv
18. ExecutiveBiz. "5 US Navy Submarine Programs Driving Undersea Warfare" (CAMP FY27 / Submarine Combat System Improvement). Aug 2026. https://www.executivebiz.com/articles/navy-submarines-uuv-ssbn-ssnx-acoustic-camp-fy27-budget

Oversight record (GAO / CRS)

19. U.S. Government Accountability Office. "Extra Large Unmanned Undersea Vehicle: Navy Needs to Employ Better Management Practices…" GAO-22-105974, Sep 2022. https://www.gao.gov/products/gao-22-105974
20. USNI News. "GAO: Navy's XLUUV Undersea Minelayer $242M Over Budget, 3 Years Behind Schedule." 28 Sep 2022. https://news.usni.org/2022/09/28/gao-navys-xluuv-undersea-minelayer-242m-over-budget-3-years-behind-schedule
21. Breaking Defense. "After $885 million, GAO warns it's 'unclear' if Navy's major UUV program will become program of record." Jun 2025. https://breakingdefense.com/2025/06/after-885-million-gao-warns-its-unclear-if-navys-major-uuv-program-will-become-program-of-record/
22. USNI News / Congressional Research Service. "Navy Large Unmanned Surface and Undersea Vehicles: Background and Issues for Congress." 25 Mar 2025. https://news.usni.org/2025/03/27/report-to-congresson-navy-large-unmanned-surface-and-undersea-vehicles
23. 19FortyFive. "The U.S. Navy Just Committed to 16 Robot Submarines…" 9 Jul 2026. https://www.19fortyfive.com/2026/07/the-u-s-navy-just-committed-to-16-robot-submarines-built-to-lay-mines…/

Peer-competitor context (claims — read critically)

24. Interesting Engineering. "Next-gen AI may end era of invisible submarines, Chinese experts claim." 14 Sep 2025. https://interestingengineering.com/military/next-gen-ai-end-invisible-submarines
25. 19FortyFive. "China's New Underwater Drones Could Blindside the U.S. Navy." 8 Dec 2025. https://www.19fortyfive.com/2025/12/chinas-new-underwater-drones-could-blindside-the-u-s-navy/
26. Army Recognition. "China deploys ships and oceanic sensors to prepare for submarine warfare against the US Navy" (Reuters ocean-mapping reporting). 30 Mar 2026. https://www.armyrecognition.com/news/navy-news/2026/china-deploys-42-ships-and-hundreds-of-oceanic-sensors-to-prepare-for-submarine-warfare-against-the-us-navy
27. Interesting Engineering. "China showcases robotic military dogs alongside anti-mine underwater UAVs" (Chengdu expo, tube-launched AUVs). 10 May 2026. https://interestingengineering.com/military/china-submarine-launched-anti-mine-vehicles

Editorial notes: (1) Chinese ASW performance figures (e.g., the ~95 percent claim) derive from a single peer-reviewed modeling study and vendor/state-media displays; they are unverified operationally and are presented as claims, not established capability. (2) The "legal" dimension of this subject is acquisition oversight — GAO and CRS reporting — rather than litigation; there is no courtroom filing specific to the edge-AI acoustics question. (3) Much U.S. tactical ASW processing is classified; open sources indicate direction and intent, not fielded performance. (4) Phonak/Sonova performance figures are manufacturer field-study claims, cited here only to establish the commercial SWaP baseline, not as defense benchmarks.