My Whisper AI trial vs. Oticon More

The day after Thanksgiving I spent most of the day talking alone to my sister in the same echoey kitchen where the loud, cacophonous family gathering had taken place the previous evening. I wore the Whispers for most of the day, switching between dynamic and custom (noise), and changing the volume up and down.

In that situation the mid-rangey tone of the Whispers did not sound that good. I was able to hear my sister’s words pretty well, but her voice sounded a bit unnatural and there was just a bit of the effect where if someone drops a pan into a sink it sounds like an explosion.

Toward the end of the afternoon I switched to the Mores. I liked the general sound of the Mores more than that of the Whispers. The Mores had none of the artificial sound of the Whispers. I thought my word recognition with the Mores in that situation was at least as good, and maybe a bit better, than the Whispers. The Mores sounded more “comfortable “ to wear.

In the evening I attended a smaller gathering at a different sibling’s in-laws’ house. I left the Mores on. This later gathering was six people in a much-less echoey environment. The Mores performed well in this situation. I could hear everyone’s words pretty well.

I stayed the night there. In the morning I started with the Whispers. Again the Whispers sounded artificial. I felt my word recognition with the Whispers was not quite as good as it had been with the Mores the previous night. I went for a hike with my brother and his wife. The Mores sounded good outside.

My takeaway from this weekend is that, so far, I like the Mores more than the Whispers, EXCEPT for that cacophonous Thanksgiving gathering where the Whispers were much better. I’m glad I still have another four weeks of the trial left.

So…Am I going to get more accustomed to the Whispers’ mid-rangey sound? Could the Whisper audi make them sound more natural, like the Mores? Would adjusting the Whispers to sound more natural reduce the SIN performance? Why do the Whispers sound so midrangey and artificial when at least some of Whisper’s people were recruited from Oticon? Will the next Whisper software update improve or fix this artificial sound? Are the Mores adjusted too extremely? I doubt I will get answers to any of these questions about the Whispers before the trial ends.

A person’s speech-in-noise comprehension can be worse than would be expected judging from their pure-tone thresholds.

[@ziploc: BTW - I’m slightly trainable, and have now learned its new, modern connotation. “Boxy” is now cool witt me!]

@ziploc said:

Why do the Whispers sound so midrangey and artificial when at least some of Whisper’s people were recruited from Oticon? Will the next Whisper software update improve or fix this artificial sound? Are the Mores adjusted too extremely?

If the fitter didn’t perform REM then your Whispers’ output could be off.

@x475aws said:

This is where you lose me. You invest more resources upfront to more fully characterize the sounds in the millions of sound scenes, and train a DNN with it. Wouldn’t you then need more resources, not less, to analyze a live sound scene in the detail needed to make use of the info in the DNN? The fact that you’re undoubtedly sampling the live scene at higher than the Nyquist frequency has nothing to do with it.

Exactly. There’s a flaw in Volusiano’s reasoning.

@Volusiano said:

Whisper decides to skin the cat the hard way, while the others used an easier way to skin the cat. I gave several examples, but I guess you don’t want to hear it or understand it. The other spent their resources and development cycle to train their DNN up front then condensed it into a formula-like result then implement this result much more easily into the HAs. Whisper meanwhile chooses to do it the hard way and that’s why they need the kind of power and resources that require a brain to implement and execute.

@happymach said:

I don’t think this analysis of the differences between the two systems is correct. From my own reading of the Whisper documentation, I’ve taken away that both brands have DNNs that are trained “up front.” What is different from about Whisper’s design is that the resulting signal-processing model is updated with every firmware cycle (every 3 months), perhaps on the basis of user-recorded data, whereas Oticon’s is more or less set in silicon and unchangeable. In addition, Whisper’s signal processing model itself appears to involve more complex (i.e more CPU intensive) real-time treatment of sound, involving more “points” of fit.

I’ve already given a detailed explanation of my interpretation of how Whisper implements their DNN vs how Oticon implements their DNN. I don’t care to repeat it here, but I’ll give a link below so anyone interested can click on it to read up on it again. I’m not going to argue the details again here, we can agree to disagree on whether my analysis is correct or not. Other readers will be the judge of what they read.

But I’ll make a few observations below:

It’s common sense to deduce that a young, small start-up like Whisper probably doesn’t have the time and resources to fully and thoroughly train their DNN with enough data using a development platform offline that can be as big as to anyone’s heart content (a la Oticon More 12 million sound scenes worth of training that probably took years to collect and process). So they take a different approach and make the brain their “online” development platform (vs the Oticon offline development platform), invest to build a library worth of a few thousands hours of unique sounds (based on what I read from their whitepaper), but not collect tens of millions of sound scenes up front like Oticon did, then continue to collect data from users via the brain storage to continue to train their DNN.

So while Oticon did all the HUGE amount of training up front so they can afford to have a finalized version set in silicon, Whisper doesn’t have the resources to take this approach, so they choose a “continuous” approach of starting out with less training data to begin with, but keep on collecting more and more users data over time to perfect their DNN. So Oticon more or less have their DNN learn as much up front then “graduate” into the silicon, while Whisper has no resources to do this so they develop a system of “learning as they go along” then eventually graduate after so many years later with eventually enough users data collected.

So while Whisper spins all this stuff about the brain being a positive thing of being able to make major breakthrough improvement updates every 3 months and continually being able to collect real life user data; to me it seems more like it’s a beta version that is released prematurely and therefore needing update every 3 months to fix bugs and introduce more lacking basic features, instead of introducing major breakthrough every few months. And the user data collection bit seems more to me like a DNN not fully and thoroughly trained enough up front, so the only way to make up for their lack of resources to train and graduate up front is to continually collect data and train as you go.

User Review of Whisper Hearing Aids

I think speech detection is not that big of a challenge. Speech has very unique characteristics that can easily be differentiated. Modern HAs from the big 6 can detect what is speech and what is non-speech very easily. But speech detection is not the challenge. Speech and noise separation IS the challenge. In the bold part I highlighted above, Dr. Schum is right that separating the noise from the speech has been very hard (almost the holy grail) to crack. That’s why beam forming to the fr…

@x475aws said:

This is where you lose me. You invest more resources upfront to more fully characterize the sounds in the millions of sound scenes, and train a DNN with it. Wouldn’t you then need more resources, not less, to analyze a live sound scene in the detail needed to make use of the info in the DNN? The fact that you’re undoubtedly sampling the live scene at higher than the Nyquist frequency has nothing to do with it.

@happymach said:

Exactly. There’s a flaw in Volusiano’s reasoning.

Not really. But I guess you guys have no clue how the DNN works so I’m not sure if I’d bother arguing with you here about it. The whole point of training the DNN to almost perfection up front is so that you don’t need as much resources to achieve good result during execution.

I’ll give another analogy. A tennis player, he spends years to train his DNN to ingrain into his brain on how to make the perfect execution when he plays. When he goes to play on a tournament, you see him make amazing, unbelievable shots EFFORTLESSLY. He does NOT need to invest more RESOURCE to execute those perfect shots. He does it effortlessly because his DNN is already trained to do it gracefully and beautifully with minimal effort and minimal resource.

@happymach said:

I still don’t think you’re right about this. And to use your analogy, as I and others have noted, the Oticon platform does not, in sound terms, make “amazing, unbelievable shots EFFORTLESSLY.”

But regardless (and it’s anecdotal), some people find the Whisper implementation to be more effective. The Oticon sound is natural, to be sure. But in terms of making speech intelligible in a noisy background, it leaves a lot to be desired (IMHO).

By the way, what makes you the expert on how DNNs work? Do you have a degree in computer science, particularly in AI?

Did I say that I’m an expert on DNN? You said that I have a flaw in my reasoning about DNN. Then I said back that you don’t know DNN the way I understand it. So tit for tat. That doesn’t make me an expert on DNN any more or less than it makes you an expert in DNN. We just agree to disagree.

When someone makes an analogy, you don’t take the literal saying in that analogy and apply it back to the original case. Of course the More doesn’t make any “shots”, regardless of whether it’s amazing or unbelievable or effortlessly.

And no, I don’t have a degree in computer science, particularly in AI. I only have an MS degree in Electrical Engineering. But one doesn’t need to have a degree in computer science and particularly in AI to talk about DNN in layman’s term, does one?

@Volusiano said:

I’ll give another analogy. A tennis player, he spends years to train his DNN to ingrain into his brain on how to make the perfect execution when he plays. When he goes to play on a tournament, you see him make amazing, unbelievable shots EFFORTLESSLY. He does NOT need to invest more RESOURCE to execute those perfect shots. He does it effortlessly because his DNN is already trained to do it gracefully and beautifully with minimal effort and minimal resource.

Thank you for posting this. It’s a beautiful analogy, but incomplete, and it tells me where your understanding of machine learning in hearing aids goes wrong. I look forward to explaining why it’s wrong, though I may end up using more tortured analogies while doing so. But I want to leave this alone for a while, so @ziploc isn’t drowned out in his own thread.

@ziploc said:

So…Am I going to get more accustomed to the Whispers’ mid-rangey sound? Could the Whisper audi make them sound more natural, like the Mores? Would adjusting the Whispers to sound more natural reduce the SIN performance? Why do the Whispers sound so midrangey and artificial when at least some of Whisper’s people were recruited from Oticon? Will the next Whisper software update improve or fix this artificial sound? Are the Mores adjusted too extremely? I doubt I will get answers to any of these questions about the Whispers before the trial ends.

There was a thread (link below, at the end of post #40 from Neville) where the OP was making comments that the More has a more natural sound than the Phonak P90, and a debate ensued, with some argument that with the right HCP who knows what they’re doing, the HCP should be able to make adjustment to the HA to make it sound like another HA of another brand/model. Neville (one of our 2 frequent HCPs on this form who are very respected) said that give him a quiet hour, and he bets that he can make the More sound like the Phonak Pico Forte, but just the basic sound, not counting things like the automatic features or performance in noise.

So maybe there’s hope that your Whisper HCP can try to tune the Whisper to sound more like the More for you, if that’s what you want. There’s nothing to lose but to ask the Whisper HCP to try for you. But I had a debate with Neville about whether you want to keep copying the sound quality of another HA brand and model forever or not? It just doesn’t seem sustainable if it has to take a quiet hour each time there’s a major update that erases the previous work and requires the matching to be redone.

Oticon More1 vs Phonak P90 - | PROS and CONS | [2021/11]

I’ve told this story here somewhere before, but I had a patient trialing a phonak hearing aid and a widex hearing aid. He preferred the phonak hearing aid sound, but liked the widex app. I matched the hearing aids in testbox. He noticed the loss of the adaptive frequency lowering, but it was close enough and he happily took the widex hearing aid. Actually, now that I’m thinking about it I had another patient who tried two sets of hearing aids and reported liking the one in quiet situations and …

Apologies if this has been dealt with and I missed it (lots of verbiage to skim through), but @ziploc’s boxy sound is very fixable is it not? The audi has all the levers they need.

@Volusiano said:

When someone makes an analogy, you don’t take the literal saying in that analogy and apply it back to the original case. Of course the More doesn’t make any “shots”, regardless of whether it’s amazing or unbelievable or effortlessly.

And no, I don’t have a degree in computer science, particularly in AI. I only have an MS degree in Electrical Engineering. But one doesn’t need to have a degree in computer science and particularly in AI to talk about DNN in layman’s term, does one?

I support @Volusiano’s responses here. His thoughts tends to be the most logically sound ones and he also admits faults when others point them out. I have a degree in AI and this analogy is elegant, although it does not capture AI performance nuances such as precision recall and other metrics when the hearing aids companies do not publicize them.

All in all, I am supportive of Whisper and hope they do well to collectively boost AI innovation in this space, which is inherently based on silicon and battery consumption tradeoffs. I would not mind making an iPhone a brain if possible perhaps in the long term.

@d_Wooluf said:

Apologies if this has been dealt with and I missed it (lots of verbiage to skim through), but @ziploc’s boxy sound is very fixable is it not? The audi has all the levers they need.

Now I don’t claim to be an HCP or audiology expert, so take what I say with a grain of salt or whatever you believe in. I only say it based on my layman’s knowledge gained as a user and what I read on this forum and on audiology papers.

The sound signature of a brand is most likely shaped through its proprietary fitting rationale. If one wants to shape the Whisper sound to sound like the Oticon sound, I assume that the HCP would probably need the Oticon Genie 2 software, and he/she would need to run the patient’s audiogram through Genie 2 to see what the prescribed gain target curve looks like. Then when he/she does REM on Whisper for the patient, instead of following the Whisper gain curve target, he/she should make the REM adjustment to the Oticon gain curve target instead. I think that would achieve the goal of acquiring the Oticon sound signature.

Now @Neville was saying that give him a quiet hour or so, he can copy the sound of one HA brand/model to another brand/model. That sounds more complicated than just mimicking the target curve of the desired sound signature. So I don’t really know beyond my educated guess above. He’s copied here so he can chime in.

Do note, however, that whenever there’s a need to re-prescribe, whether it be due to a change in dome fitting, or change in audiogram, or perhaps an in-situ audiometry run, then the duplication process needs to be redone again, both on the Genie 2 side to generate a newly prescribed target curve, then on the Whisper side to adjust the REM results toward the Oticon target curve instead of the Whisper target curve.

So in theory I think it can be done. How practical it is depends on how often it needs to be done due to updates and changes and whether re-prescriptions need to be performed upon updates and changes or not. It may also not be perfect if the fitting dome on the Oticon side may not be exactly the same compared to the fitting dome on the Whisper side, so any adjustment on the Oticon prescription for the new fitting may not fit perfectly with the fitting on the Whisper side, unless you use the Oticon domes on the Whisper ear pieces.

@Volusiano said:

The sound signature of a brand is most likely shaped through its proprietary fitting rationale.

Ok. My view might be simplistic. I was thinking @ziploc needs more of the high frequencies, so just give him more of the high frequencies. Maybe the fitter was being very conservative in that respect, possibly treating him more as a new user? I don’t know what tools they have to work with. I wasn’t thinking of making it sound like anything else, actually. Just give him more highs.