Hacker News Reader

This comment is technically incorrect.

Noise-cancelling headphones work the way OP described, and are limited to working in a cavity which is smaller than the wavelength.

There are predictive noise cancelling systems too. These are commonly built into industrial equipment. You might, for example, have a fan or motor which spins at 5000 RPMs, and what do you know, it has a consistent noise profile on each of those rounds. If you cancel that noise profile, you do pretty well. These systems both reduce noise and extend equipment life by reducing vibration. They also work to higher frequencies (in part because speed-of-sound in metals is > 10x that in air, and what matters is wavelength more than frequency).

There are ones which are open-loop used in HVAC systems. If I have a pipe, I can record sound on one end, and play a cancelling waveform on the other end, with an appropriate delay. I haven't followed the field, and when I looked, these were mostly expensive prototype systems (I'm not sure if they ever made it into mainstream use), but they did work. They needed to be calibrated for each HVAC, which made them impractical for most real-world applications.

There are all sorts of ways to do active noise cancellation. Many of these could be implemented on airplanes. Most of these techniques were implemented decades ago, when computation was a lot more expensive. A simple feedback loop is a few opamps, capacitors, and inductors, so pretty cheap to toss into headphones with 1990's-era technology. One could do a lot more with teraflops....

Yes, as I said, this doesn't work in the general case when your noise source is uncorrelated and not predictable. If you have a predictable noise source then you've solved the time travel problem and you can cancel it any way you want as long as you can obtain feedback from the location of the user's ears.

We're talking headrest ANC here (for use cases like consumer ANC headphones), not built-in noise reduction for industrial systems.

An airplane is an industrial system....

And in either case, the design of algorithms depends on much more on context than on whether it's "consumer" or "industrial." In this case, the context looks a lot more like industrial machinery.

The actual study is here [0] and the abstract and introduction do a much better job of describing what the study is actually about.

I encourage you to read at least those before you dismiss the research and its real-world applications.

TL;DR: The purpose of the research is to try out a new way of figuring out which sound is going into the user's ears, so that it can be cancelled out. Current methods work for lower frequencies but not for higher frequencies.

They "cheated" because they're testing one component of the system and the "cheating" component is immaterial to the research they're doing.

[0]: https://www.nature.com/articles/s41598-020-77614-w

They "cheated" because they did the easy part, and the "other component" has no known solution.

What they're handwaving away is the real problem. The reason why current methods work for lower frequencies and not for higher frequencies is that your microphone, emitter, and ear all need to be enclosed within a volume << the minimum wavelength you can handle, to eliminate sound field and propagation time effects and let you treat the noise cancellation problem as a point problem in space.

Once you put the ears in open air and the emitters away from them, the problem isn't figuring out the sound that's going into the users' ears, it's figuring it out in advance. The ANC speakers have to send the complement signal before the original signal arrives at the ears, so they both arrive coincidentally.

In other words, the "missing part" to make this research work is time travel (if you use their sensing technology as the only input). Or maybe full-volume sound field sampling and characterization technology (if you don't). Neither of which exist, and which are orders of magnitude more difficult than this research.

The handwave part is "The constraint of the possible locations [of microphones] would affect the control performance and this remains a topic to be further studied in the ANC community."

Yes, it would completely destroy high frequency performance, which is what they were trying to achieve, in anything but lab scenarios (such as the sound sources still being loudspeakers and the microphones being directly in line between them and the user). Real world noise doesn't come from loudspeakers, it comes from all around you. Good luck using distant microphones to compute the expected sound at a user's ear in advance, with any kind of accuracy in the high frequencies.

At best I imagine they could achieve adaptive cancellation for a set of slowly moving pointlike sources in an otherwise simple room, with a number of microphones greater than the number of sources. It's like RF MIMO systems. But again, this is orders of magnitude more work to implement, and real world constraints are going to kill your high frequency response. And for larger sources - forget it. You just can't characterize the transform for that. Not enough dimensions in your input data to solve for it. So anything mechanical, stuff where the noise isn't coming from a literal speaker with a 1-dimensional input signal - nope. As the uncertainty and source size grows, your high frequency response goes down the drain.

No, the missing part is not time travel. Sound comes from the skin and engines of the plane, where microphones could be located, and could potentially predict sound throughout the plane well in advance with an appropriate microphone array. You'd need a fair number of microphones, but I suspect the noise profile isn't too chaotic, which helps. You're talking about things like jet engines (which are repetitive), turbulence (which seems chaotic, but in practice, you shed vertices at a known rate), and a known cabin configuration (modulo people moving about). I suspect while the general solutions suffers from the curse of dimensionality, the specific one probably doesn't. A reasonable number of microphones, combined with decent models, could characterize the whole sound field.

Speakers, located in the seats, if they knew where ears were, could play sound to cancel that.

Decent microphones are $2, and what was expensive were things like microphone preamps, ADC, and the whole data pipeline. From there, we need a shit-ton of computation which really wasn't practical until... big-ass computation systems came out for the rise of ML.

The problem is hard today, but definitely not impossible. I don't think I would have given that same answer a decade ago. An NVidia Titan V brings over 100 teraflops. That's a lot of cycles one can throw at trying to predict sound by my ear from sound at the skin of the airplane in real-time.

If you've ever been in an airplane, you'll know that airplane cabin noise is largely uncorrelated rumble, not fixed-pitch components from the jet engines and such. Good luck solving for the response at a user's position for that, no matter how many microphones you throw at the problem.

https://www.youtube.com/watch?v=te3pyLk_wBs

Spectrogram sample: https://mrcn.st/t/flight_noise.png

I get that this is HN, but no, ML and AI do not magically solve all problems.

Thankfully, most of the energy is in the low frequencies, which existing noise cancellation systems can already do a god job of dealing with in the near field.

Research is an incremental process and even if their proposal has no practical application right at this very second, it's still a new idea that hadn't been tried before they did it.

Instead of focusing on the fact that these researchers didn't singlehandly revolutionise ANC in a single study, how about we focus on what they did do?

That's why I'm confused as to why they're selling this as being relevant for ANC (optimizing for the news cycle?), when it would be a lot more useful to talk about things like optimizing free space speaker systems for a user's position, say, for VR.

Like, the idea of remotely sensing where a user's ears are and what they're picking up is useful. Just not for far field ANC.

> This works for wavelengths smaller than the cavity size

(A) "This" works for a wavelength < L.

> and above a certain frequency all you have is passive muffling

(B) "This" does not work for a frequency > F. <=> (B') "This" works for a frequency < F.

---

Don't A and B' (or B) contradict each other? Because an upper limit for the wavelength equates to a lower limit for the frequency - not an upper limit.

What is your background that justifies your very confident demeanour?

It's clearly a typo. Should be "This works for wavelengths larger than the cavity size."

A long wavelength implies that the phase of the reference input and feedback input are similar, which is what allows you to just subtract them.

Your processing delay, the distance between the speaker and the microphone, and the speed of sound are all fixed, so as the wavelength gets shorter, the relative phase difference between the two signals you're trying to subtract increases. At a certain point, you run out of phase margin and can no longer just subtract them anymore.

If you were designing an ANR system, you'd probably want to put a low pass filter somewhere around this point to prevent your feedback system from oscillating.

Maybe that's why the speakers of Sony's xm series are designed to lie flat on (my) ears while Bose's leave more space but are considered slightly less efficient at noise cancelling compared to the xm's. For me ANC is by far the most important feature but having my ears pressed for longer than one hour is painful. And I tend to wear my Boses (QC35 and HP700) easily 8+ hours.

Sorry, that was an edit typo. I originally had "frequency" but I figured that would be confusing when juxtaposed to "cavity size", took the inverse and forgot to flip the sign.

My background is audio processing is one of my hobbies, I have torn down and repaired noise canceling headphones and understand how they work (and have previously talked about this on Reddit and been corroborated by an actual engineer at a company designing such headphones), I have implemented laser galvanometer closed loop feedback systems and worked with them (which is both a similar control system problem, and a different part of OP's research paper on the light path), I have attempted room and speaker response cancelation (and wrote my own DSP firmware for implementing biquad filters, which I use for my DIY living room rig), making an open source speaker calibration system is in my future project list and I've recently done a lot of thinking about this... and, most importantly, since this is neither my professional nor academic field I have no reason to overstate my claims or mislead about the importance of my work, unlike way too many academic researchers (see also: that guy who I recently called out for being a paper mill about side channel leakage methods after he successfully spun a false media story about generating Wi-Fi signals from RAM using a misleading paper title; in that case infosec is in fact one of my professional fields).

I don't like to brag or anything, I'm just a curious guy who likes tinkering with (lots of) things... but if you insist on asking "what are your credentials", well, there you go.

So, in your opinion there's no hope for active voice cancelling? Maybe with earbuds?

I'm a silence junkie living in world getting noisier with every day it seems - so this subject is of high personal importance to me. That's why I'm interested in getting an idea about where to place your assessment.

Active noise canceling earbuds already exist, and probably perform about as well as one can hope modulo small incremental improvements. I mean, the AirPods Pro are already damn impressive here. You could probably do better by cramming more and more technology closer to the actual ear /at the tip of the earbuds, perhaps? These things do cancel the low-mid end of voice frequencies quite well. I imagine with enough miniaturization you could increase the range.

But room-scale or free-space noise cancellation only works when either 1) your noise source is extremely predictable, or 2) your noise source is low-dimensional. It's just never going to work for the general case unless you can somehow sense the entire soundfield. The entire sum of what waves are traveling in which directions. Not just a few microphone inputs. It's going from 1D data to 3D data, the difference between a light sensor and a CAT scanner. And it needs to work at 20kHz response.

I guess hypothetically you could do something like wrap a room in microphones and compute the sound coming in in all directions, then with a very good adaptive model be able to cancel noise from outside within it. But if you're going to draw a boundary and wrap it in a massive array of microphones... Aren't you better off just adding insulation material? :-)

Edit: just as a baseline, I very crudely tested my Bose QC20 earbuds (which aren't exactly bleeding edge tech) and I think they get about 10dB active cancellation at 1kHz, which goes down to nothing at 2kHz. Human voice fundamental frequencies are typically 85-180 Hz for adult cis men and 165 - 255 Hz for adult cis women, so this level of ANC does get rid of the fundamental and a few harmonics. I might test a friend's AirPods Pro later and see if they're better. 1kHz has a wavelength of ~34cm, so I think there is room for improvement here in the earbuds case. I'm not experienced enough in this field to have a good feel for the numbers beyond order of magnitude estimates, but I think going up to 4kHz might be doable, given that for earbuds we're talking about cavities in the ~5cm range. Might need more mics/drivers and more miniaturization to pull it off, not sure.

For those wondering, the test methodology was to play a tone, turn on cancellation (which has a delay), then turn off cancellation (which is instant) simultaneously with decreasing the amplitude of the tone, and trying to match the perceived loudness between both cases.

I'm more of a lazy pragmatist than a passionate hacker. So, I guess I should invest my energy, time and money into just finding a less noisy place to live. I'm not suffering from misophony - I just hate hearing my neighbours or loud vehicles with a vengeance :D

See, this is why when I shop for apartments to rent, concrete walls are a requirement. Not just so I don't hear my neighbors, but also so they don't hear me, since I make music ;)

At my current place I've pretty much concluded that anything but deep bass just doesn't transfer. Even at 5 in the morning, I can listen to stuff at a reasonable listening volume without disturbing my neighbor, as long as I'm not pumping out bass below 80Hz or so. And the neighbor recently got a dog and I haven't heard her bark even once when she's home :)

> But if you're going to draw a boundary and wrap it in a massive array of microphones... Aren't you better off just adding insulation material? :-)

Well, let's do the math here. I can't find the data sheet for the weight of an electret microphone cartridge, but let's assume 4 grams. A thousand microphones -- assuming we want 100 the length of the airplane and 10 around -- is around 4 kilos. Add another 2 kilos for computation, and a few more for wiring, and you've added the weight of a piece of luggage.

Now, let's say you want to add 2 inches of mineral wool sound insulation. I'm assuming 100 meters x 10 meter circumference (I have no idea, but the numbers scale the same). You've added (quite literally) half a metric ton to your airplane. And made it 2" thicker, either reducing cabin space or increasing drag.

Plus, sound insulation does very little for low frequencies, which is the majority of airplane noise.

Assuming you want to cancel frequencies up to 1kHz, you're going to need mics every 30cm at least for a useful sound field. For a 787 fuselage, that's about 60 around by 190, let's round that down to 10000 microphones. But of course, then you need electronics for them, and a way to carry the signals back to somewhere for processing. That's about 10 Gbps of data, fine, doable, but then you need to process it. Since you need to cancel separately per passenger, that's 2.4 million filters (10k mics * 242 passengers) to process. Nevermind that since this is a plane, there is a huge amount of overhead due to safety requirements.

This is all assuming this whole idea works, which it probably won't, because it's not just about the microphones on the fuselage but also how sound is transmitted inside the plane and other noise sources.

That sounds reasonable. Let's do the math your way in both directions:

1) Assuming system-level model

10,000 microphones * 8 kilosamples per second = 800 megasamples per second.

NVidia Tesla does 100 teraflops. I get about 125,000 FLOPS per sample. That feels adequate to me!

2) One filter per microphone per passenger. We need to divide by 250.

500 FLOPS per sample per passenger. That's more than enough for a very fancy IIR.

Of the two, I think #1 is more likely to work than #2, precisely because you want a coherent model. If you want to adjust your model for someone walking down the aisle (or any kind of system ID), that's a lot easier with 125,000 FLOPS per sample than 500 FLOPS per sample.

Right, and now you've achieved the same noise cancelling performance... That off the shelf noise canceling headphones achieve (1 kHz) without putting 10000 microphones on the plane.

Again assuming this all works, which is a massive if.

Look, this isn't practical.

.... in the same way that having a billion transistors in a computer isn't practical. Or a computer in every car tweaking fuel injection. It's not practical until it is. Electronics goes down in price. Algorithms improve. Insulation stays fixed or goes up in price. I would say it's not a question of "if" but "when."

The answer to that might be right now (we couldn't do it before, and we probably can today) or it might be in a decade. But electronics will keep falling in price. 10,000 microphones * $5 per microphone+electronics = $50k.

The advantages of cancelling an entire wavefront go well beyond passenger comfort too. Industrial noise cancelling systems are more about equipment life than about employee comfort. I had laptop screws unscrew on airplanes before, due to vibration. If planes need less maintenance as a result of active vibration reduction throughout the airplane, you'll make up that $50k virtually overnight.

As a footnote, the crazy part here isn't the 10,000 microphones, but the speaker-at-every-seat part. You'd almost certainly want both the microphones and speakers in the skin of the airplane. But that's a story for another day.

Some cars such as new Genesis models attempt this.