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.
1 Like
Not so fast. You do claim some authority, by saying things like “I guess you guys have no clue how the DNN works,” which is actually quite different from stating “you don’t know DNN the way I understand it.”
When you make statements like, “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,” you’re also claiming authority beyond what could be qualified as “layman’s.”
However, I agree that there definitely is an aspect to the Whisper that smacks of testing – and that’s normal for a completely new, just-released hearing aid like this. I think that soliciting customer feedback, regardless of whether you judge the product to be immature or not, is a good thing – by comparison, Oticon only deals with audiologists. The Whisper is certainly not for everyone, but I wouldn’t make the judgment that it was released “prematurely” – that’s just the nature of software/hardware development which, especially in the initial stages, will never be close to perfect.
FWIW, I think that you get irritated or angry reactions to your posts about the Whisper because, as pointed out earlier, there is a generally dismissive, skeptical tone that you adopt vis-à-vis the product that is annoying because you have no personal experience with it.
However, I’m happy to agree to disagree about Whisper’s merits.
1 Like
Not much more complicated. If ear canal coupling acoustics are nearly identical, it’s easier to do it in the testbox. You just run the first set of hearing aids, and then adjust the gain and compression and max output on the second set of hearing aids to match exactly. You need an REM system that will run multiple versions of the same stimuli or, better, one that saves comparison curves. If the coupling is significantly different, you need to do it on-ear which is annoying because the patient needs to sit still and quiet for a while.
But no, you don’t typically need to do it again later. When you update the audiogram it doesn’t re-set everything. Similarly, if you make a new manual program you just duplicate the base program that is set up and go from there. I said a quiet hour to make a More sound like a pico forte, but a pico forte is an ancient analogue BTE–going from one modern aid to another modern aid is faster. (I was probably over-estimating with an hour anyway.)
I’ve never seen a whisper though.
In the original Whisper thread, when nobody had trialed it yet, and where anything goes, yeah sure, I spoke my piece about what I think and was dismissive of the key ideas of Whisper (the brain, the subscription model, the pricing, the no DIY).
But in this thread, I challenge you to find anything I said that’s being dismissive of Whisper, NOT UP UNTIL THE POINT when @x475aws asked me whether I felt an ethical obligation to disseminate the information from Whisper’s whitepaper about it being able to process the size and capability of the deep learning algorithm like nobody else can. Of course I don’t, and I listed out the reasons why, it being that I dismiss its claimed approach as the ONLY legitimate approach that everybody has to measure up to.
So the door was opened by @x475aws inviting me in to enlighten me on the “new” information, and wanting me to accept it and asking if I don’t feel an ethical obligation to disseminate this “new” information as an influencer. That’s only when I objected and became dismissive, meaning I spoke my mind, again. I really don’t want to beat this horse to death again, but it kept on being resurrected and waved in front of me, wanting me to submit to it.
And the questions asked did not relate to, nor require that I need to have any personal experience with it to form my own opinion. It never asked for my opinion BASED ON my personal experience with Whisper. It only asked for my opinion BASED ON the “new” information presented.
Just for reference, below is the quote on the question from @x475aws posed to me in post #94 on this thread that opened the door to my (dismissive) response.
1 Like
Exactly, MrV … Exactly!
You just need to quit defending yourself. There are, always have been, and always will be it seems, those whose self esteem seems to be wrapped up in, and insisting on the universal application of the choices they make. You can’t change them and there’s a point where simply ignoring them is best. I’ve found the things you’ve said to be helpful and anyone who can read without prejudice will benefit from your input whether or not the product you prefer works for them or not. You’re not the one who looks ridiculous. In the words of Mark Twain, never argue too long with fool for those looking on might not be able to tell which is which.
5 Likes
No, that’s not it. It’s that Whisper is a true breakthrough, from my experience and that of a few others here, and I want it to survive. The hostility to Whisper, among some well-spoken people here, is certainly not helping its chances. If it survives, its work will eventually benefit all of us, because all the manufacturers will have to offer something like it. It might become the top tier of aids, and push down the price of what is now the top tier. Whisper has hearing industry business people who are certainly aiming for insurance and VA coverage. But if it doesn’t get traction, and the venture capitalists pull the plug, then I’m afraid we won’t see this kind of speech-in-noise performance (along with the most open-ness) again for a long time, for any amount of money.
2 Likes
I may have misidentified the motivation but the result is the same. Not every question, not every disagreement, not every doubt is illegitimate and if a product is worth its claims it will survive. You’re just a bit too gung ho and quite out of line in suggesting that Volusiano has a responsibility to support your (as yet unproven) claims of Whisper superiority. I’ve read the testimonies here and the best ones don’t include your “this is the greatest ever” assertions. Maybe it is and time and performance will tell, not the anecdotal experiences of a few here on this forum.
Additionally I don’t see any hostility on the part of Volusiano. His hesitancy/refusal to jump on the Whisper bandwagon based on what is yet limited evidence is hardly hostile. When I explained that the More 1 was not working well for me he suggested that I might want to check out the Whisper as some here were experiencing some very positive benefits. That can hardly be called hostility. The best thing you can do is show how its positive benefits have worked for you and leave it at that. There are so many variables at work in individual hearing issues that there is most likely never going to be a one size fits all solution.
5 Likes
Taking it from there, say our tennis player shows up with his leg in a cast. Even with his trained neural network, he isn’t likely to win that day unless it’s a really friendly match. Or he shows up with a visual handicap of some sort, missing super high diopter contacts or retinal detachment or whatever. He isn’t going to play up to his usual standards. Or he’s short of breath because of asthma or cardiovascular issues. He won’t be able to keep up.
Yes, our tennis player knows how to play tennis, after many hours of practice and many games. But he still needs bodily resources (sensory, musculoskeletal, respiratory, circulation) to get around the court and deal with the ball in every moment of every game.
Should I try to make the jump back to DNN’s and hearing aids, to finish explaining why processor speed matters even when there’s a trained DNN?
Please do! You’ve lost me!
In the analogy, the DNN is the brain. Resources are the bodily functions. At least by your post above, you agree that resrouces are bodily functions. There’s no argument that the DNN is the brain, I hope.
My argument is this: simply that take 2 men, same basic bodily functions, and one man trains his brain to play tennis and the other doesn’t train his brain to play tennis. The tennis guy will win in the tennis game. No extra resources like super bionic $6M dollar man stuff needed to be installed in his body to win in the game of tennis, in case he breaks his leg, or gets short of breath, or scrapes his eyes out.
So the Whisper argument is that just because there’s a brain, extra resources is needed for that brain to work. I disagree. The brain just needs to be trained properly. The development effort is in the training up front to be able command the body efficiently and avoid injury. That’s the whole point of using the brain, because you want to avoid and don’t want to have, nor do you need, nor do you want to pay for the extra cost and baggage for the bionic stuff. You simply want to train the brain so you can use it with your mere mortal body.
Perhaps if somebody has to Frankenstein the heck out of his body in order to support what the brain can command on that bionic body, then maybe that brain has not been trained properly the way it should be, to be able to just be used on the mere mortal body.
OK, so maybe Frankenstein wins the game of tennis over the tennis player. I don’t know about you, I just want to play tennis well enough that I’m happy with the game of tennis in my mere mortal body. I don’t want to turn into Frankenstein in order to beat everyone at tennis. To me, there’s an elegant component to tennis that can’t be beat by a brute.
Exactly, @Volusiano! Exactly!!
I’m not sure that x45aws understands the nature and function of analogies.
Ha ha, I was going to reply graciously to the initial version of your post, and then decided I’d better wait for the edits. Glad I did.
I figured that if we came to an agreement on what I thought was a straightforward example of a human tennis player, where the “DNN” is the biological neural network in his brain, then I would try to craft an apt analogy to transition to a DNN hearing aid. It wouldn’t be easy, because the analogy isn’t exact, and analogies lead to more heat than light in my experience. But if you’re saying that the DNN in your tennis player example represents the Whisper Brain, then you’ve really lost me.
What a great saying!
I’m not sure if Volusiano intends this, or if it is just the way my mind has modeled it. I view in my mind a model where the DNN is burned into silicon. It learnED at one point. It has learned. But it will no longer learn, it reacts to what it is given as inputs in light of what it learned and this is reflected in the outputs which change how other parts of the HA are configured at that moment. The silicon implementation is blindingly fast and efficient power-wise. I’ve seen AI/ DNN implemented in a similar fashion for other applications. Not saying this is how it is or must be implemented in any HA. I am ignorant there.
To Volusiano’s argument, should we call it that, if the DNN is burned and fixed, more horse power of processors isn’t going to help a whole lot. If you somehow implement the DNN in general purpose silicon, or graphical processor type implementations, that will consume huge resources and not give you wonderful response times.
I haven’t worked in systems using these technologies for some years, so I’m out of it. Does anyone know how they implemented this stuff?
That’s the WHOLE thesis of the Whisper technology, its “Sound Separation Engine, which is based on an AI model” (the DNN), their words, not mine, “is enabled by the processing capability of the Whisper brain”, again, their words, not mine.
In other words, if Whisper didn’t have the Sound Separation Engine DNN, then they wouldn’t have needed the power of the brain to process it. The question is “why do they need to processing power when the others don’t?”. The answer is “because they implement their DNN differently than others.”
The question is “Which way is better?”, the answer is “We don’t know because nobody wants to share much information anyway.” Like I said earlier, there are many different ways to skin a cat.
But it’s obvious that in one way, you need a lot of processing power but maybe less training effort up front. In another way, you invest in a lot of training effort up front so you don’t need a lot of processing power at implementation time. It’s consistent with how a young startup company would do it vs how a much bigger old guard would do it.
1 Like
Do I detect a fellow Electrical Engineer here in @whitehat? 
Yes, your statements in bold above are pretty much what I intended. But I’d go further another step to say that the direct to silicon implementation without the need for further processing power is ALSO possible and made easier because what is burned into silicon are simply optimized weight and bias values of the neurons in the DNN networks, which are already calculated and optimized from the off line training development effort. So now it’s a matter of storing those hard coded values into silicon ROMs for lightning fast calculations to propagate the calculated data from the input to the output of the neural network in no time.
Such a model will give lightning fast results, while a more general purpose model (maybe a graphical processor type like you mentioned) would require a lot of processing and calculation and memories to arrive at the result because it has to “figure” things out in real time. With the earlier (faster) model, the “figuring out” is already done up front through the off line training development work.
1 Like
@Volusiano: Yeah … Oticon doing all that homework up front is sorta like playing tennis with Chuck Norris. Chuck is sooo fast he never serves the ball - you call ahead for take out.
Sorta like that with Whisper, eh?
Imagine Chuck Norris playing tennis against John McEnroe -> Chuck is so fast he never serves the ball -> John McEnroe goes “YOU CANNOT BE SERIOUS!!!”
1 Like