Industries / Health / Medical devices / Hearing aids
Fitting
Fitting is the work of measuring a person's hearing, choosing a hearing aid and setting its amplification for that hearing, then checking the result in the ear.
7
Milestones achieved, from 1922 to 2022.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1922
Hearing measured by machine
An audiometer plays tones of known pitch and loudness through an earphone, so hearing can be measured frequency by frequency. Western Electric's 1-A audiometer, designed by E. P. Fowler and R. L. Wegel, tested from 32 to 16,384 Hz and cost $1,500, "slightly less than a house"; 25 were sold in 1922. The 2-A of 1923 was smaller, cheaper and ran on dry cells, and became the first widely used audiometer, by Wayne Staab's account.
Lessonthe early machines had no standard calibration, and the first ANSI standard method for manual pure-tone audiometry came only in 1978, Staab notes.
Credit
E. P. Fowler and R. L. Wegel; Western Electric
1946
Aids chosen by speech tests
In 1943 the US Army opened four aural rehabilitation centers to issue hearing aids to servicemen deafened in the war. At Deshon General Hospital in Butler, Pennsylvania, Raymond Carhart's team worked with more than 16,000 of them and settled on a "hearing aid evaluation": try several aids and keep the one with the best speech-test scores. Carhart set it out in "Tests for Selection of Hearing Aids" in December 1946. Earlier that year the "Harvard Report", by Hallowell Davis and colleagues at Harvard's Psycho-Acoustic Laboratory, had argued that fitting aids one person at a time was not justified.
Lessononly two of Carhart's four tests survived into civilian clinics, yet his comparison method dominated hearing aid fitting for the next five decades, James Jerger wrote in 2018.
Credit
Raymond Carhart; Hallowell Davis and colleagues
1986
Gain set from the audiogram
A prescription formula turns a hearing test into target amplification at each frequency, in place of trying aids one after another. Denis Byrne and Harvey Dillon of Australia's National Acoustic Laboratories (NAL), founded in 1947 to serve children and war veterans, published a revised formula, NAL-R, derived from measurements of 30 ears. Tested on 44 people (67 aided ears) by paired comparisons, its prescribed response was "seldom inferior to, and usually better than" the variations tried.
Lessonthe paper says it was needed because research showed NAL's previous procedure did not meet its own aim of making every frequency band of speech equally loud.
Credit
Denis Byrne and Harvey Dillon, National Acoustic Laboratories
2011
Prescriptions for compression aids
Modern aids use wide dynamic range compression (WDRC): they amplify soft sounds more than loud ones, so a prescription must set gain for several input levels. NAL-NL2, published in 2011, aims for the most intelligible speech at comfortable loudness and adjusts for the wearer's age, sex, experience and language and for compressor speed. Its authors built in field data: in 187 adults, women preferred about 2 dB less gain than men, and new users with moderate or worse loss preferred less gain than experienced ones. The rival DSL v5.0, from the University of Western Ontario (2005), extended the Desired Sensation Level method to infants, children and adults with compression aids. In a 2018 survey of 251 audiologists, 51% set gain with a prescription formula and 35% with the maker's own first fit.
pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
Gitte Keidser, Harvey Dillon and colleagues (NAL); Susan Scollie, Richard Seewald and colleagues (Western Ontario)
2018
Fittings often go unchecked
Real-ear measurement puts a thin probe microphone in the ear canal next to the hearing aid's tip to check that the sound delivered matches the prescription. US guidelines recommend it, but Michael Valente and colleagues wrote that 70 to 80% of audiologists do not use it routinely and rely on the maker's first fit. In their double-blind crossover trial of 24 new users, aids set to NAL-NL2 with real-ear checks gave a median 15% more words correct for soft speech (50 dB SPL) than the same aids on the maker's first fit, and 79% of participants preferred them. In 2026, Kruger and colleagues compared the first fits of 29 flagship prescription aids from 11 makers, released 2021 to 2025, with NAL-NL2 targets on an acoustic manikin and found no improvement from older to newer generations.
Lessona first fit is the maker's estimate; only a measurement shows what reaches the eardrum.
Credit
Michael Valente and colleagues; Kruger, Manchaiah, Taddei, Bailey and Swanepoel
2018
Fitted by the user
On 5 October 2018 the US Food and Drug Administration granted Bose's De Novo request, creating a new device type, the "self-fitting air-conduction hearing aid" (21 CFR 874.3325, Class II). The Bose Hearing Aid was for adults with perceived mild to moderate loss and needed no hearing test or pre-programming: the user set it with two controls in a phone app, one for loudness and one for fine tuning. In Bose's study of 75 adults, the benefit of self-fitting on a speech-in-noise test (QuickSIN) and two questionnaires was not inferior to an audiologist's fitting, and self-fitters preferred their own settings.
Lessonthe evidence was the company's own trial, and every participant first had an audiologist choose their ear tips, so the study did not test choosing tips alone.
Credit
Bose Corporation; US Food and Drug Administration
2022
Fitting from 70 miles away
In synchronous tele-audiology the audiologist programs the aids over a live video and data link while the patient sits elsewhere. In the Conexiones trial, 28 mostly low-income Hispanic and Latino adults aged 55 to 89 were fitted at a rural health center in Nogales, Arizona, by a clinician at a university 70 miles away, with a trained community health worker or a student beside them. Both groups gained in listening self-efficacy, and the authors judged the model feasible.
Credit
Laura Coco, Nicole Marrone and colleagues, University of Arizona, with Mariposa Community Health Center
Watch for
NAL-NL3 replaces NAL-NL2
NAL says its NAL-NL3 Prescription, built from 1.33 million adult fitting records, a survey of 76 clinicians and a deep neural network trained on the optimised targets, predicts higher intelligibility than NAL-NL2 at normal loudness. It gives less gain below 0.5 kHz and above 6 kHz and a little more from 1 to 4 kHz, and 29 clinicians who used it in 263 fittings reported easier fitting and fewer fine-tuning changes. A Comfort in Noise module, tested on 83 users, lowers gain in noise.
Achieved whena group with no tie to NAL publishes a blinded comparison of NAL-NL3 and NAL-NL2 fittings on measured speech scores or preference.
Claimed by
National Acoustic Laboratories, papers published April to October 2026
Open
01
First fits across many audiograms
A public table of how far each maker's first fit lands from prescription targets across all ten standard audiograms of Bisgaard, Vlaming and Dahlquist (2010), from mild flat to severe sloping loss. Program current flagship aids from at least five makers to each audiogram with the maker's default first fit, measure them on an acoustic manikin or in a test box with recorded speech at 55, 65 and 75 dB SPL, and report the error from NAL-NL2 targets at each frequency, adding NAL-NL3 when fitting software offers it.
The briefCredit · unclaimed02
Fine-tuning advice, scored
A public test of AI models as fine-tuning assistants. It holds 400 written cases, each a standard audiogram, the aid's current settings and a complaint in a wearer's words ("my own voice sounds hollow", "it whistles when I hug someone"), with an answer key giving the change experts agree on. Score general AI models and small models fine-tuned for the job against the keys, and count both correct and unsafe answers.
The briefCredit · unclaimed
Open 01 · the brief
First fits across many audiograms
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of how far each maker's first fit lands from prescription targets across all ten standard audiograms of Bisgaard, Vlaming and Dahlquist (2010), from mild flat to severe sloping loss. Program current flagship aids from at least five makers to each audiogram with the maker's default first fit, measure them on an acoustic manikin or in a test box with recorded speech at 55, 65 and 75 dB SPL, and report the error from NAL-NL2 targets at each frequency, adding NAL-NL3 when fitting software offers it.
Why it is still open
The largest outside study, Kruger and colleagues (2026), measured 29 flagship aids but used one audiogram, the moderate sloping N3, as HearAdvisor's lab does for every device. Sanders, Stoody, Weber and Mueller (Hearing Review, 2015) measured the premier aids of five makers in 16 real ears, all programmed to one sloping audiogram. Valente's 2018 trial tested one maker's first fit. HearAdvisor's matched retest of two aids in October 2026 found first-fit gains 3 to 4 dB apart on the same audiogram, enough to move its speech scores.
pmc.ncbi.nlm.nih.govhearadvisor.comhearadvisor.comhearingreview.com
Built with what exists
An acoustic manikin such as GRAS's KEMAR, or a hearing aid test box with a 2cc coupler (a small cavity standing in for the ear canal) such as the Audioscan Verifit 2; the makers' fitting software, used through a university audiology clinic; the ten standard audiograms (Bisgaard et al.); the International Speech Test Signal; Python to compute root-mean-square error.
First step
Program two makers' flagship aids to the N2 and S2 audiograms, measure them in a test box, and set the results beside the published N3 data.
Done looks like
At least 5 makers × 10 audiograms × 3 speech levels, each fit's root-mean-square error from target published with the raw curves, and a second lab repeating three of the fits within 2 dB.
Open 02 · the brief
Fine-tuning advice, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI models as fine-tuning assistants. It holds 400 written cases, each a standard audiogram, the aid's current settings and a complaint in a wearer's words ("my own voice sounds hollow", "it whistles when I hug someone"), with an answer key giving the change experts agree on. Score general AI models and small models fine-tuned for the job against the keys, and count both correct and unsafe answers.
Why it is still open
The answer keys exist on paper. Jenstad, Van Tasell and Ewert (2003) turned the 40 complaint words patients use most into troubleshooting steps on which 24 fitting experts showed "a high degree of agreement", and Thielemans and colleagues (2017) built a Dutch version, independent of any maker, from 112 dispensers and 15 experts. The AI studies so far test other things. Batuk and Karakuluk-Celebi (2026) scored ChatGPT and Gemini on 44 frequently asked questions about hearing aids. AUDIOLOGYBENCH (Li, Fitzgerald, Wang and colleagues, Journal of Medical Internet Research, 2026) graded four frontier models on 67 clinical cases that ask for audiogram readings, diagnoses and recommendations, not for changes to an aid's settings, and found critical errors in 35.4% of the models' audiogram readings. The CAFA prototype (Ding and colleagues, arXiv, September 2025) turns complaints into tuning commands without a published scored test set. Signia's app assistant is described only by its maker's staff.
europepmc.orgeuropepmc.orgeuropepmc.orgarxiv.orgpmc.ncbi.nlm.nih.govegms.de
Built with what exists
The 2003 and 2017 expert tables; the ten standard audiograms; AI models such as Claude, GPT and Gemini through their APIs; a small open model fine-tuned on one graphics chip, such as Liquid AI's LFM3 (see the Note "One GPU to fine-tune"); three licensed audiologists to review the keys; Python for scoring.
First step
Write the 40 cases implied by Jenstad's 40 complaint words and run three AI models on them.
Done looks like
At least 400 cases with keys reviewed by three audiologists and their agreement stated; for each model, the share of answers that match the experts' direction of change and the share that propose an unsafe change such as raising maximum output, all published with the cases.