6 ms·

This is a really cool project. The one issue I see is that the Teensy isn't what you'd call power efficient (compared to some more domain-specific DSPs, that is

This is a really cool project. The one issue I see is that the Teensy isn't what you'd call power efficient (compared to some more domain-specific DSPs, that is). >That seemed incredibly expensive. I haven’t done much electronics, but digital signal processing doesn’t seem that expensive? The cost is in R&D, not just for human trials but for DSP engineer salaries. It's an extremely specialized skillset and the companies that hire for it look for PhDs most of the time. In particular, optimizing adaptive filtering to clock in the dozens of cycles per buffer with fixed point DSPs to maximize battery life, while providing features like feedback reduction, noise cancellation, working around the infamous "cocktail party" problem, with modern features like Bluetooth... all with audio fidelity on par with some serious audiophile gear. My understanding is that there are also some other market forces too like very long lifecycle management and low volumes for new products, but idk how real that is. Point is, I don't think those devices are really overpriced - at least no more than any other medical device paid for by insurance companies. EDIT: I'm reading some of their documents and a few things jump out at me. They mention supporting FFT/IFFTs on the platform. This stands out to me since doing that on a battery powered device is unusual. Firstly because it's not cheap in cycles, and secondly because useful FFTs require rather large buffers of memory which harms your latency, and may come at a premium. When I've seen them required (e.g, a codec), it was typically on a dedicated FPGA with the algorithm burned in, not running on the DSP.

terminalhealth 7y ago
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?
- holy_city 7y ago
  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.
- azalemeth 7y ago
  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.
jdietrich 7y ago
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 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.
- jacquesm 7y ago
  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.
  - rrss 7y ago
    > 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.
    - shae 7y ago
      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.
      - throwaway9298 7y ago
        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.
rrss 7y ago
> 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.
- cushychicken 7y ago
  Plus, many ffts actually benefit from fewer points and less bandwidth - especially in voice applications, where much of the spectral content is under 4kHz.

terminalhealth 7y ago

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?

holy_city 7y ago
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.
azalemeth 7y ago
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.

jdietrich 7y ago

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 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.

jacquesm 7y ago
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.
- rrss 7y ago
  > 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.
  - shae 7y ago
    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.
    - throwaway9298 7y ago
      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.

rrss 7y ago

> 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.

cushychicken 7y ago
Plus, many ffts actually benefit from fewer points and less bandwidth - especially in voice applications, where much of the spectral content is under 4kHz.