Device that can hear your “inner voice” | Hacker News Reader

Device that can hear your “inner voice”

theguardian.com

22 pointsPaulAJ14 comments

Full paper can be found here [1], and the MIT press release with a much more reasonable title here [2].

I've been wondering for some time how one would make a practical DIY electromyography (EMG) device, for example to measure hand muscle activation around the wrist, but I haven't been able to find a source that presents research in an accessible manner. Does anyone know of such a source?

[1] https://dam-prod.media.mit.edu/x/2018/03/23/p43-kapur_BRjFwE...

[2] https://news.mit.edu/2018/computer-system-transcribes-words-...

How DIY are you trying to build? Basically you just need a set of electrodes, an amplifier, and a data acquisition board. Getting good data is a black art, though that's true with many different things beyond EMG.

The only way these types of devices will ever gain traction is if they are unobtrusive or subdermal. And subdermal is the upgrade path—no one will volunteer to have surgery in order to try something out. This is doomed unless they can figure out how to make it unobtrusive. No one's going to wear a janky-looking jawbone, even by themselves.

Well, almost all forms of technology start off being ugly.

Ugly is fatal for wearable tech.

Smartwatches still sell, 90% of them are (were?) obtrusive, bulky pieces of junk.

You don't wear smartwatches on your face.

Back in 1990 the novel "Earth" by David Brin predicted this.

He has snuck that tech into a couple of his other books too, though my recollection is that it got more “inches” in Earth.

I find it frustrating when articles start out by misrepresenting the technology with "that can read people’s minds". It's understandable why they simplify things like that, but as someone who was involved with general EEG work and LFP work in the speech-motor cortex it feels like an unfair comparison. That's not to say that observing muscle movement isn't pretty rad, but they're different technologies reading different types of biosignals.

Can you provide more perspective on this & how well it can likely recognize subvocalized words?

I'm wondering if someone won't find a way to do this remotely. I'd like to think that impossible, but they managed to do some very odd things like finding heart rates from pictures, so I can't really say that they couldn't somehow find a way to pick up the signal remotely and I'm curious how good the current tech is from someone who as actually used it.

Sure. Looking (quite) quickly at the paper there's a few things that stand out.

- Small vocabulary sizes (20 or less words at each decision level)

- A lack of discussion as to identifying utterance onset

- A lack of trials of the same subject over many different days (EEG/EMG/LFP can vary considerably across subjects and within the same subject in different trials)

- A lack of discussion on the cross-validation protocol (the setup of training and testing folds can provide radically different views on computational neuroscience sorts of results)

All of this is fairly normal within this sort of work, however it also means that a hefty grain of salt should be taken when viewing the work. With the results that they have you have 90% accuracy when picking 1 out of 5-20 words if you setup the system and trained it the same day. That's pretty cool for small muscle movements, though not really a solution that's directly usable IMO. (also don't believe the results that are claimed by [6].)

> a way to do this remotely.

i.e. with a standard camera at a standard resolution/framerate? There's coarse information that you can likely see from another human yourself if you see someone mutter something, but there's too much to muddle things for it to be useful immediately. If you get to the point where you're using high resolution high framerate multi-spectral cameras you might be able to get marginally more information, but nothing too major IMO. The MIT heartrate enhancement worked well due to the periodic nature of heart rates as well as the proximity of blood vessels to the surface of the skin. Neither of those help out for observing a-periodic muscle movement under a lot of skin.

This was incredibly informative and I just wanted to say thank you for providing a lot of perspective on this because a single upvote wasn't enough.

This is basically the missing piece that enables early mass adoption of AR. Think about it, no matter how small they make it people aren't going to go outside in public wearing their Magic leap or hololens glasses and speaking out loud to an alexa or Cortana or Siri or waving their arms in the air like some possessed lunatic. This however is game changing. I can sit on the bus and speak, in my head to my virtual assistant and have it control the virtual part of my augmented reality. This will turn AR from something I'm comfortable doing only in the privacy of my own home to wearing it 24/7 everywhere I go. The future is going to be wierd and wonderful and even more isolated.

Reply on news.ycombinator.com