Skip to content
HearingTracker
Log in

Fortell AI 2.0: Next-level Speech-in-Noise Performance Update for Hearing Aids

9 replies527 viewsLast reply

Link to this topic
Reply
5 people in this conversation5 min readJump to latest

BTW, we have tested the aids in the lab. Results to be published mid Sept.

Thanks for letting us know, I'm looking forward to the results.

Looking at the two published papers for this product, the multi-manufacturer comparison is limited by the use of preference rankings at a 0dB in the absence of any speech recognition testing. It tells us that the 20 adults tested preferred the Fortell spatial but doesn't tell us exactly why this was the case or that it resulted in better speech recognition performance vs the competing products. The choice of a 0dB SNR is also not necessarily useful in the real world as many listeners are performing at chance levels at that SNR and are giving up listening anyway. An SNR80% where participants are scoring 80% of the words correct is a more useful level of understanding for comparisons to be made.

The study where speech recognition is being compared using a single manufacturer comparison and it isn't possible to figure out which product the comparison device is from the information provided. It is possible that poorest performing device with an onboard DNN was chosen rather than the best in class device. The use of the using fixed SNRs for testing isn't ideal as there is no normative data for this testing format and a lot of uncertainties about the reliability and validity of test when administered this way needs to be addressed. It would've been better to use validated tests that are designed for use with this format like the connected speech test or AZBio sentences.

It should be noted that the study parameters were designed to maximise the effects of the spatial noise reduction with used and features such as , wind noise reduction and transient noise reduction disabled. These features would typically be enabled for a real fitting, so the real world performance of Fortell's AI denoiser is likely to be less than that measured in the studies.

Many forms of feedback cancellation would have the ability to significantly interfere with a spatially-based AI noise reduction algorithm as implemented by Fortell, particularly those used by Sonova and Starkey which use acoustic tagging via phase changes in the sound outputted by the hearing aid as a means of reducing feedback risk by giving the aid the ability to distinguish sound already amplified by the device from environmental sound. These phase changes are of the type that would be used to determine the original direction a sound came from that Fortell's design relies upon.

Edited

Thanks for the detailed engagement! To respond to the points you raised:

1). We study intelligibility and subjective preference separately because we feel each alone is insufficient. We think a hearing aid must: restore audibility, improve signal to noise ratio sufficient to enable understanding, and keep voices sounding natural.

2) For the subjective study which compares Fortell to all 5 top of the line competitors, the output of the control hearing aids averaged 0dB but ranged from -5 to +5 , thus spanning a reasonable range where hearing aid wearers might attempt but struggle with conversation. By contrast, in the same scenes, the Fortell device was able to produce an average output of +8 dB, which explains the overwhelming preference for Fortell.

3) The comparison device used in the intelligibility study was the market leader for -based hearing aids. At the time of writing (still today AFAIK) it was the only hearing aid on the market employing a DNN for true source separation (as opposed to scene classification). Even though we did not disclose the manufacturer (so as not to be disrespectful to peer manufacturers) they were outed by an independent journalist at Wired magazine (https://www.wired.com/story/hearing-aid-startup-ai-fortell/10).

4). , wind noise reduction, and transient noise reduction were NOT disabled for any of the human testing in either study. They were left on in their standard default settings on all devices in all studies. The only place where these features were disabled were on the lab measurements that use the Hagerman-Olofsson SNR measurement technique. Turning off these features is necessary and standard practice for this technique because it requires unperturbed measurement of phase.

Hi Matt, thanks for your response.

@MattHdeJonge said:

At the time of writing (still today AFAIK) it was the only hearing aid on the market employing a DNN for true source separation (as opposed to scene classification).

I'm not sure what you mean by "source separation" here. Other manufacturers like Oticon, Resound and Widex use their for noise reduction, not merely scene classification and have done so from their first product release.

@MattHdeJonge said:

the output SNR of the control hearing aids averaged 0dB but ranged from -5 to +5 dB, thus spanning a reasonable range where hearing aid wearers might attempt but struggle with conversation.

If it averaged out at zero then the level was already below where many hearing aid users are performing at chance levels. It'd be good to see you test at SRT80% correct in the future as this is a more realistic situation. If you can achieve a superior where users scores 80% correct then that is solid proof of a superior product that has meaningful real world benefit. If you're a few SNR better at getting chance level performance it's still chance level performance. We need the other part of the performance-intensity curve.

Edited

We will definitely continue to study our product in a variety of environments and SnRs and will keep this community updated as we have more data to share. Likewise, HearAdvisor will soon publish data. For now, you might reference Figure 3 in our intelligibility paper. Here we plot %-words-understood vs . You can see that all 30 blinded participants showed materially improved listening accuracy. SNR-50 moved by over 9 db vs the state of the art "" competitor. Even SNR-80 showed material improvement.

Thanks, I read the figure described but it seemed to be based upon mathematical extrapolation rather than actual data. We'd really like actual data using tests that are designed to be administered at fixed .

In reply to @Hearing_Potential

@Hearing_Potential said:

I'm not sure what you mean by "source separation" here. Other manufacturers like Oticon, Resound and Widex use their DNN for noise reduction, not merely scene classification and have done so from their first product release.

Simply Infinio Sphere. Except Fortell it was the only product performing meaningful source separation similar to, e. g. Adobe Podcast . According to one of the Dr Cliff's videos in Aug 2024, Spheric can be effective even with one microphone in hearing aid.

In reply to @Hearing_Potential

My apologies: above I referred to figure 3 (corrected). Figure 4B shows you the raw data without any extrapolation (presented or required). You can see how participants achieving -80 with the control device performed when switching to Fortell. Eyeballing: they went from 80% to 95% understanding. You are correct in that we did perform Quick Sin at 3 SnRs: -6, 0, +6 db in this study.

Related discussions