34 comments

[ 4.3 ms ] story [ 84.0 ms ] thread
Sounds like anyone wanting to do a "proper" fake job will now need to separate out any existing hum, then add a new continuous "hum track" based upon those publicly available ENV values.

Probably wouldn't even be all that hard.

By itself? No. But this is just one more thing that can be used to date a recording. If you even know about it.

There are hundreds of known fingerprints and you just need to forget about or not know of just one for your fraud to be called out.

Well subtracting the hum requires knowing the hum at that time, right?
Yes,*and* its harmonics, which would have to be replaced as well if they're detectably present. Swapping hums isn't trivial.
If the goal is just to obfuscate the hum, couldn’t you just scramble it instead of “subtracting”? Especially if you’re not trying to replace the hum
Wouldn't something like Nvidia's noise cancellation tech (as commonly used by podcasters, etc) remove the original hum well enough?

eg ready for new "fake continuos hum" to then be applied

A video on the same subject: https://youtu.be/e0elNU0iOMY

I really enjoy Tom Scott. Everything he puts out is wonderfully concise.

I like the guy too but not a big fan of the whole dumb-it-down/ELI5 culture. Some subjects absolutely require a lot more (written text, experimentation/tweaking with code) than few fancy videos to grasp.
All subjects have non-dumbed down versions in research papers, university level books or industry reference manuals. The existence of more dumbed down content doesn't make those disappear.
A dumbed down version is great to introduce people to the subject

Chances are that good written explanations already exist, just a Google search away, for those who want to learn more.

Tom’s video and this blog post are perfect combo that complement each other.

Dumbed down without being inaccurate is also often sufficient for anyone with some base knowledge in an area. You can figure a lot out simply by being told the core concept, no need to go into extreme detail.

Extreme detail is great for those that want or need it, but it's far from the only reasonable option for learning things.

---

Re the general flood of these kinds of videos: oh yeah, I think it's excessive. I suspect YouTube is pushing that category lately (which is FAR from the worst to push, so I'm largely okay with that). It feels like all the popular, algorithmically pushed ones are making sure to cross the magic 10 minute mark, and try to look like TV/movie editing, regardless of what those do to the content. Bonus points for obtuse video titles that tell you nothing about what topic is covered.

Which is part of why I like his. They almost never hit 10 minutes, or even close. There's little fluff, little drama, little algorithmic gaming - videos today look very similar to several years ago.

IIRC the reason they aim for the 10 minute mark is because they can put in mid roll ads, not because the algorithm wants it.
It's actually often impossible to dumb something down without inaccuracies. I'm always drawn to the metaphor of resolution in an image. If I want to simplify an image by reducing its resolution, I'll have 4 pixels at a time that I need to decide how to represent with 1. Whatever solution I do will ultimately result in an image that is technically a less correct version that the original image.

If I start a class by saying "computer code will run line by line on your processor, one instruction set at a time" this might be a useful way to get someone familiar with writing code. If someone raises their hand and says "well don't CPUs perform optimizations to try to do branch predictions other optimizations so that it doesn't execute exactly as written, and isn't some code able to run on GPUs, and can't cosmic rays cause bit flips in memory so that your code doesn't execute exactly as written, and..." Sure, but I'm teaching you a wrong thing because it's functionally right at this resolution.

Have you tried any 3Blue1Brown videos? Highly recommend
Yes and I believe their videos are one the best things on the Internet right now as far as math education is concerned. Same goes for {Computer,Number}phile.

The reason they're high quality learning material is because the explanations of the various concepts respect weaker math backgrounds, without being desperate to be as reductive as possible or have clickbaity titles , and sometimes with downright wrong claims, to generate more audience.

An example of a channel that consistently follows this latter pattern is Veritasium i.e. see the electricity video controversy.

Agreed on all points, and yeah. Veritasium was more alright a while back, but they've gone WAY off the deep end nowadays. Beyond merely "no longer interested in their content", I think they might be actively damaging at this point.
His company accounts would suggest he's popular.
Yes, Tom Scott is one talented YouTuber.

See, for example, his video on microwaving frozen hamsters back to life[0], in which he interviews the 101-year-old J. E. Lovelock, the scientist who did the work back in the fifties.

[0]https://www.youtube.com/watch?v=2tdiKTSdE9Y

RIP James E. Lovelock, who passed away at the age of 103 in 26 July 2022.

I can see the passion in his tones in the video.

Great article. Something in me finds satisfaction when I recognised something I learned about at uni applied in the real world.

Now on to building an ENF spoofing machine that people can carry around with them to confuse eavesdroppers.

Curious if this can be used to discern AI-generated voice recordings vs authentic.

A determined faker could always add their own noise, but this be a good first-pass detector.

A recent approach along these lines is to do acoustic modeling of the vocal tract that would have produced the speech. Unsurprisingly, AI voices do not correspond to realistic human vocal tracts, for now.

https://www.usenix.org/conference/usenixsecurity22/presentat...

can google's cloud measure our vocal tract topology from the bluetooth and wifi signals on our phone? expect this flaw in AI voices to be self healing, retrained, and corrected soon.
So yet another argument for proper, balanced audio recordings and working noise gates.
(comment deleted)
Wouldn't help. The electrical noise will still be present when the noise gate opens (because the thing we're recording is above the threshold), so they can extract the frequencies they need from parts of the recording when there's signal (e.g. the voice we're recording). And since the signals can come from what the microphone captures, it's irrelevant if the cables are balanced.
Right below the author's comment, "Regardless of a grid’s size, the current flowing through it oscillates in perfect synchrony, at exactly the same frequency, throughout the whole grid," he presents a map of Western Europe's regional groups. I notice that Finland, Sweden, Norway, and Iceland comprise the Nordic Regional Group.

Iceland? There is a lot of water between Iceland and the other three nations in its group. Is Iceland's power indeed synchronized with the other three? If so, why does it need to be?

Below the map, the author includes a link crediting the Wikipedia article on power grids[0], which includes maps of other regions in the world.

Perusing the map for North America, I notice another noncontiguous regional group, namely the Alaska group. Most of Alaska is off the grid (or, more accurately, on a collection of local grids too small to be named). The named group includes a slice of the interior extending northward from Anchorage towards Fairbanks and a separate noncontiguous area containing Juneau. I have the same question, are the two sub-regions synchronized?

[0]https://en.wikipedia.org/wiki/Wide_area_synchronous_grid

(comment deleted)
The map for Europe is a map of organizational/bureaucratic boundaries within the ENTSO-E; the world map on that page correctly omits regions not connected to the big synchronized grids.

Alaska, similarly, is a set of two isolated grids, connected only by organizational ties: https://en.wikipedia.org/wiki/Alaska_Interconnection

Norway to Faroe Isles and Faroe to Iceland are about the same distance individually as the electrical links between the UK and Norway and UK and Denmark, so I don't see why not. The north atlantic is a little deeper, but that's probably not much of an increase in distance all considered - a couple of km at each end at most.
> Regardless of a grid’s size, the current flowing through it oscillates in perfect synchrony, at exactly the same frequency, throughout the whole grid

I have a feeling that this is not exactly true. The grid is composed of transmission lines (in the “wires on pylons sense”), that are on one hand short enough that they do not really behave like transmission lines (in the circuit analysis sense), but on the other hand long enough that their propagation delay and other “RF-like” properties are somewhat interesting for design of efficient grid.

On the other hand, the phase shifts and related distortion effects are outside of the bandwidth of the audio recording as long as the grid is at least somewhat properly designed.

I'm wondering how difficult it is to clean the signal once the relevant noise portion can be extracted, and then introduce a fake noise to manipulate the date?

I'd wager it's not too difficult to produce something that can fool a layman just using some dating software...