Who has better AI, Fortell or Phonak?

Difference 1: how sealed the ear tip is

For the hearing loss we test with, a standard moderate sloping audiogram (N3; Bisgaard et al., 2010), Fortell prescribes a power dome. Phonak prescribes a more vented coupling. A more occluding tip blocks more of the direct, unprocessed sound that would otherwise leak past the device into the ear canal. That matters for noise reduction. The leaked sound carries the original noise, so the less that gets through, the more of what you hear is the cleaned-up signal.

We measure this directly as real-ear occluded insertion gain (REOIG): how much the ear tip alone, with the device powered off, reduces sound at the eardrum compared with an open ear. From 1 to 4 kHz, Fortell's tip blocks about 18–21 dB. EON's vented coupling blocks almost nothing at 1 kHz and about 10–14 dB at 2–4 kHz. When we fit EON with a power dome, it lands in Fortell's range above about 1.5 kHz, blocks a few dB less around 1 kHz, and seals considerably more below about 700 Hz.

The cost of occlusion shows up in our own-voice score, which penalizes the “head in a barrel” sound of your own voice when the ear is sealed. Fortell scores 1.5. EON scores 2.8 with its vented coupling and 1.1–1.3 with a power dome.

Difference 2: the first fit

Out of the box, the two devices apply very different gain to the same audiogram. The chart below starts at a 75 dB speech input, the level closest to our Speech in Loud Environments scenes (71–80 dB SPL). There, EON's first fit tracks the NAL-NL2 prescription closely up to 2 kHz but falls about 5 dB short at 4 kHz. Fortell's first fit runs about 4 dB above target at 250–500 Hz and at 2 kHz. The net result is that Fortell delivers roughly 3–4 dB more gain than EON at 250 Hz and from 2 to 4 kHz, the frequencies that carry most of the consonant information in speech. Our speech metrics are sensitive to gain, so part of the published gap is simply a gain difference.

The experiment: match gain and occlusion

To isolate AI de-noising processing, we re-measured EON with a power dome, which matches Fortell's occlusion, and tuned it to the NAL-NL2 targets, which matches gain. At a 75 dB input, the two tuned fits agree within about 1.5 dB from 250 Hz to 4 kHz, and both sit within 2 dB of the NAL-NL2 target (the matched panel above). At 55 and 65 dB they agree within about 2 dB from 500 Hz to 4 kHz.

We score speech in noise with two metrics. HASPI v2 predicts objective intelligibility: how many words you would get right. The HearAdvisor model is our proprietary model of subjective ease of understanding, trained on more than 100,000 quality-controlled listener ratings. It predicts how hard the listener feels they are working. Our published score blends the two 50/50.

Matching gain and occlusion shrinks the difference dramatically. In Speech in Loud Environments, the gap shrinks to 0.2 when both are matched. On the HearAdvisor model, subjective ease of listening, the two are equal (5.6 vs 5.6). On HASPI, objective intelligibility, Fortell keeps a slight edge (6.1 vs 6.3).

Scene by scene, Fortell is ahead or even in six of the seven Speech in Loud Environments scenes. EON wins only one, the dinner party, but by enough to pull the averages nearly level.

The takeaway for Speech in Loud Environments: both companies have class-leading AI with very similar performance levels, perhaps with a slight edge for Fortell.

In conversations with us, Fortell pointed out that we did not test its hyper-directional “Front Voices” mode, which is off by default (following the logic in our decision tree). We don't know how that mode would compare to Phonak. Similarly, we did not test EON with power domes and SSC engaged, so we don't know how that configuration would compare to Fortell.

In Speech in Quiet and Moderate Environments, Fortell has a real edge

In the matched condition, Fortell still leads in Speech in Quiet and Moderate Environments (below 70 dB), with a 50/50 score of 4.6 vs 3.9. Most of that lead comes from the two reverberant church scenes, where the HearAdvisor model favors Fortell by about 2.3 points. The observed differences likely come down to when each company engages its AI processing. Fortell's runs all the time, whereas Phonak's default automatic program (AutoSense OS) only switches to its AI-based “Spheric Speech in Loud Noise” program when it classifies the scene as noisy.

You can see this in the estimated latency. In Speech in Quiet and Moderate Environments, EON's end-to-end delay is about 6 ms. In Speech in Loud Environments it steps up by a flat ~7.4 ms to about 13.5 ms, which is the signature of the Sphere network switching in. Fortell measures 9.4 ms in both: the AI is always running, and it runs at a lower end-to-end latency than Phonak's does when switched on.

So who has better AI?

Fortell scores higher than Phonak EON in our published results. Only a very small portion of its advantage in loud environments is due to its AI denoising. When the gain and ear tip are matched, the two products perform very similarly in loud environments. Fortell's published noise lead comes mostly from a first fit with a more favorable gain profile for these metrics and a more occluding default dome.

Where the products differ more is when and how their AI denoising runs. Fortell's AI is on all day, in every environment, so it also helps in Speech in Quiet and Moderate Environments. Phonak's AI only switches on in loud environments. Fortell runs at a lower latency in Loud Environments: 9.4 ms, against about 13.5 ms for EON once its AI switches on. Phonak has less latency (6.1 ms) in quiet and moderate environments. Phonak also has many other advantages that HearAdvisor testing does not consider, such as Bluetooth Classic streaming (with better wireless link quality) and far richer fitting features for audiologists.

Then why does HearAdvisor test this way?

If gain and occlusion move the scores this much, why don't we always match them? Our approach has always been to measure the settings a consumer is most likely to experience, since consumers are who our website is for. So we use the manufacturer's prescribed ear tip, the one a user is most likely to wear. Similarly, the first fit is the gain profile most users will actually live with. For the same reason, we measure the first fit and weight it more heavily than the tuned fit (77% vs 23%).

This experiment has been enlightening, though. It shows that a single headline number can mix together the quality of a device's AI and the choices it makes about fit. We're wondering whether a different decision tree in our methods would be more useful to share with readers. If you have ideas, please reach out.