Hacker News Reader

Hearing aid processing chips are incredibly sophisticated and specialised - just check out the block diagram in this datasheet:

https://www.onsemi.com/pub/Collateral/SA3291-D.PDF

That chip is doing graphic equalisation, multi-band dynamic range compression, noise cancellation, feedback cancellation, data logging and wireless communication with an absolute maximum power dissipation of 50mW. It's a $100 chip for good reason.

The costs of a chip have very little to do with what is in them. It's a $100 chip for a good reason: profits. Because if you can't hear then $100 extra is very little money. Once the first silicon samples are in and the numbers are large enough it is very rare to see a chip that would command such figures in other fields, there is no reason that I am aware of that would justify this price.

> numbers are large enough

How many people have hearing aids, though? Seems like the market for this would be much smaller than other things and therefore the price higher.

While writing this blog post I've discovered that way more people have hearing problems than actually buy hearing aids. The largest part is the expense, but another part is the stigma. No one wants to wear a hearing aid if they're under the age of forty or so.

50 yo. congentitally profoundly deaf person here. Cochlear implant fitted last month. My quality of life has been utterly transformed: I can't imagine what life would have been like if I had had one from birth. If appropriate for the patient, the healthcare system should skip the hearing aid and go straight for an implant. If you happen to be reading this, and have a hearing loss, explore cochlear implants whatever your age.

Whilst I wish the OS hearing aid project every success, what we really need is open source implants - the pricing is even more insane than hearing aid technology and is relatively well understood. The existing, certified implants could be used, but with an OS/DIY processor - circumventing many of the potential certification issues.

I have a friend who is the perfect candidate for a cochlear implant, but refuses to have closed source inside his head.

He does not wish to have his body hostage to any company, and I can understand that.

Thanks for sharing your story and glad to read about your positive experience. Just a comment regarding your advice to others.

I went to bed with perfect hearing and then woke up with 90% hearing loss in one ear (sudden nerve deafness). Some of it has come back but the gain is not flat across frequencies. It is incredibly disorienting. I was almost hit by a car the other day since I now look in the opposite direction when I hear a car accelerate. I no longer enjoy listening to music and it limits my ability to be witty in casual conversation since I now often second guess what was said. Music sounds weird.

I know a hearing aid would help but I don't want anyone to know I have this problem. Hopefully my brain can just recalibrate the frequencies especially since I have one pretty normal ear.

Regarding your suggestion. I believe a good cochlear implant has about 22 channels. A healthy human ear can discern 300,000+ frequencies. This is why cochlear implants are only given to totally deaf patients. The technology is no where near where it needs to be used on patients with hearing loss.

More people would buy them if they were cheaper. That's simple economics. They are priced right now to extract maximum $ from the market, not to ensure that the largest number of people have access. 100's of millions of people world wide have hearing issues, some 15-20% iirc.

> That's simple economics.

Sure, it's simple economics, but in this case it is too simple.

It turns out that the hearing aid market has inelastic demand [1][2]. In 2010, market penetration for hearing aids in the U.S. was around 24% (8.2 million users). Amlani estimated that with a complete subsidy of hearing aids, market penetration would only increase to 34% (11.2 million users) [1].

Combined with the lifespan of 5-8 years, that is a quite small scale for an asic. (Apple will sell 40+ million airpods this year alone).

[1]: Amlani, A. M. (2010). Will government subsidies increase the US hearing aid market penetration rate? Audiology Today, 22(2), 40-46.

[2]: Lee, K. & Lotz, P. (1998). Noise and silence in the hearing instrument industry. Working Paper, Department of Industrial Economics & Strategy, Copenhagen Business School.

> Amlani estimated that with a complete subsidy of hearing aids, market penetration would only increase to 34% (11.2 million users) [1].

That's an interesting use of the word 'only', you're looking at a 30% or so increase and at that scale this would not have a huge affect on the price of the chip itself because once the costs of developing and the start-up costs have been born the rest is marginal. These chips probably cost < $2 to produce even at this quantity.

What you seem to forget is that once it is worth doing an ASIC that is pretty much proof that the economies of scale are there. The very rare cases where an ASIC is still expensive is when they are top of the line in switching speed, density or pin count and these devices have none of that.

> That's an interesting use of the word 'only'

A ~thousand to several thousand dollar product with a market penetration of 24% (low) becomes completely free, and the market penetration shifts to 34% (still low). I think 'only' is justified.

> the costs of developing and the start-up costs have been born the rest is marginal

The NRE costs have not been 'born' until the product is EOL. They are distributed across the price of each unit. My hypothesis is that this chip has a sufficiently low volume that the NRE cost / chip is substantial.

> What you seem to forget is that once it is worth doing an ASIC that is pretty much proof that the economies of scale are there

When you have strict power requirements (i.e. need a battery life of a month), it can still make sense to go for an ASIC even with relatively low volume. Add to this an inelastic demand curve (i.e. you will sell the same number independent of price), and there isn't a compelling reason to try to do it with a DSP or FPGA.

> These chips probably cost < $2 to produce even at this quantity

If we assume that OnSemi could design this chip for $10M, then they would have to sell 5M of them to have your proposed unit cost of < $2 (assuming a wafer cost of zero, which is obviously wrong). I would guess that $10M is a lowball for the total development cost, and that 5M is way optimistic for volume (that would pretty much require this random chip is in 100% of hearing aids sold in the U.S. in the last few years). They've probably sold an order of magnitude less than that.

Maybe I'm wrong here, but it isn't as obvious to me as it apparently is to you.

I think cost is a large factor because retirees have fixed income. If insurance pays for a two thousand dollar hearing aid that lasts three years, you get it. If insurance doesn't pay, you become socially isolated.

Do you know of other hearing aid options cheap enough to not require being covered by insurance?

Every single chip function cited here is available in any given wearable/hearable chipset, many of which are available in consumer products that cost less than $100. Economies of scale suggest that the price of the actual chip that does all that is less than $10. Personal experience suggests it's likely less than $5 - maybe even less than $2 in large quantities.

A big part of why hearing aids haven't faced much disruption, as I understand it, is their qualification as a medical device. A major competitor of my current employer had a policy of many years that they would not do hearing aids, in spite of extensive audio and audio processing experience internally. The reason? Amar Bose wanted nothing to do with medical device certifications.

So what you're saying is that if someone put in the effort to get through device certifications they could disrupt a rather profitable market. I feel like it would be relatively easy to get approval with a 510k (basically saying your device is similar to a pre existing device)

Yeah, I suppose I am.

Easier said than done, though.

>Every single chip function cited here is available in any given wearable/hearable chipset

All but one - low power. A typical pair of true wireless headphones runs for three hours on a charge; a typical hearing aid runs for a week on a zinc-air cell. Any Turing-complete processor will do whatever kind of DSP you need it to do, but the trick is doing it incredibly efficiently.

>All but one - low power

Even that part is becoming less true. I've seen multiple datasheets in the last year that promise worst case FFT/IFFT performance at 2mA. Measured performance has been less than a quarter of thst.

That block diagram looks like any other embedded device available today. Less complicated than most. No video processing in there, so lots less complicated.

And most are under $10.

> This stands out to me since doing that on a battery powered device is unusual

This is absolutely not true. I would bet that > 50% of all devices shipping with a digital signal processor are doing FFT/IFFT at some point.

> useful FFTs require rather large buffers of memory

Loads of realtime FIR filtering is implemented with overlap-add or overlap-save - i.e. block-based FFT/IFFT.

Plus, many ffts actually benefit from fewer points and less bandwidth - especially in voice applications, where much of the spectral content is under 4kHz.

The question is how much these additional fancy features actually improve the product over what is feasible DiY? How much longer is the battery life?

I'm no expert in the field, but it's my understanding that those features are why modern devices allow people to hear in noisy environments. Feedback reduction especially, which is not easy to deal with (tl;dr, the mic/speaker are mechanically coupled, with a poor seal and loud enough environment you get feedback and the DSP has to compensate).

And in terms of battery life, I don't know. I do know that most battery powered audio devices with DSP throw floating point math out the window from the get-go, and I haven't seen a job opening for DSP in hearing aids that didn't mention fixed point math in awhile. I don't know of any processors that fit the bill there however, those things usually have a proprietary IDE/debugger/flasher you need to pay for.

I'm reasonably sure it would be possible to implement appropriate DSP Voodoo Magic (aka finite response filters) on the Teensy that does a _pretty_ good job of all of these things.

However, I agree with the OP that it does have a fairly large number of shortcomings:

-- The number of man-years that have been spent on commercial products is huge, and they provide a highly optimised solution in a competitive domain. My late grandfather had, in the 1970s, a hearing aid with feedback reduction powered by an analogue computer (aka "electronics") that hung around his neck. -- A lot of the tuning parameters of these algorithms really are hardware specific and would require quite a lot of tuning /iteration -- At the end of the day, a Teensy is a moderately large rectangular board that will not fit behind your ear, has nonzero power requirements, and is a general purpose CPU. A 3D printed case is an expensive way of making a plastic box to put it in. If you were going to go down the open hardware route, you'd start somewhere very different, with power efficient dedicated DSP units on a small, thick multilayer board milled to be a bit more ergonomic. A modern hearing aid needs a new battery every month or so, and is powered by a 0,54 g 1.4 V 180 mAh battery (that is a 4x6 mm cylinder [h x d]). You're not going to get anything like that for a general purpose CPU.

Still, this is a fun project, and I commend people doing it. As ever with anything to do with the USA and healthcare products, however, I can't help but think that their efforts would be better spent trying to get universal healthcare. The cost to the NHS for two hearing aids, multiple fitting appointments included, is around £400.