Here's a video of Dave from YC trying it out: https://www.youtube.com/watch?v=9QdX1eObVYI
(Edit: you don't actually have to move your lips or even open your mouth for this to work! Since the RF can see inside your mouth the system can be fine tuned on "internal speech". This isn't the automatic tongue movements that occur when "thinking" words to yourself, but more like mouthing with your lips closed. It's a bit weird to do, but with practice, we have been able to get it down pretty well.)
Arnav and I were studying CS at Georgia Tech at the beginning of the year, with no experience in silent speech or sensing. We were both AI-native builders and power users of Wispr Flow. What bugged us was not being able to use these tools in the library (or an office) or in public. At Stanford’s TreeHacks in February, we decided to build the most technically ridiculous thing we could think of. We’d read about AlterEgo, a silent speech system out of MIT’s Media Lab, and wanted to recreate it in 36 hours. We built a crude prototype, sticking EMG electrodes onto my face:
https://drive.google.com/file/d/1mwhv5HZMz3UTevdrgwS5O7g2jWV...
We got it to kind of work at a letter level in a very janky way. It wasn’t good enough to win the hackathon, but convinced us to go all in. The first time I typed a sentence without anyone noticing, I felt the magic of the technology.
We quickly realized EMG (reading muscle electrical signals) wouldn’t work:
- Signal changes due to sweat
- Electrical interference
- Placement variability
- Most importantly, no one wants to stick electrodes onto their face!!
So we switched to RF and built the glasses you see now. They emit radio waves, similar to Bluetooth, that pass through your face. As they cross the tongue, lips and soft palate, they’re modulated by silent mouthing. We train foundation models on this data to decode these perturbations to text.
Despite prior academic research, getting this to work was a huge engineering challenge. First we built an extremely low noise RF sensing stack. Then we had to place antennas so their fields overlapped maximally with the articulators. Our first attempt on glasses failed completely, so we reverted to old habits and taped antennas on our face:
https://drive.google.com/file/d/1ks0dS0eZiTV9EqSFsyB_w5BBjpc...
Since taping antennas consistently in the same spot and angle was hard, we created this weird but effective head rig to hold them in place:
https://drive.google.com/file/d/1rhzNFpdDzr7oHk0isFY70dkoV4C...
This worked, but obviously wasn’t a product yet. So we worked deep in simulations to design glasses that could hold the antennas.
That evolved into our design now, with antennas embedded within the frames. Once we established there was a signal, the next hurdle was cross-user generalization. Everyone’s facial anatomy, and hence RF signal patterns, are vastly different. We tried algorithmic workarounds but realized the only lever was to massively scale pretraining data.
So that’s what we did. We set up booths in our office and paid people to read prompts wearing our glasses, to pretrain our models on. This produced a beautiful scaling curve, with errors falling quickly as we scaled our dataset and cross-user performance becoming genuinely compelling (as you see in the video above; Dave only had a short 20-sentence calibration).
We always heard about building things from first principles but never understood it before this. We tried tackling it top down but completely failed. Only when we broke each component (the antenna, electronics, ML) down to its fundamentals to understand and optimize it did we get it to work really well.
Over the next few months we’re continuing R&D to turn our prototype glasses into the slick product on our website and ship in Q1 2027. We’re also scaling up our dataset to improve performance on edge cases.
For anyone who loves voice dictation but hates using it around people, we’ve opened preorders for our first batch of 500 units, with a $200 discount off the final price. Allocation is running out quickly, so get yours at https://subvocal.co/preorder!
Second, how tough is this signal processing? Will it be done on-glasses, or through an app? Third, what's the equivalent of a wake word for these? I'm reminded of the Apple Vision Pro radar to see if your fingers are clicking. I could imagine a lot of modalities where I want to just start talking and have things happen, and a lot where I don't want the glasses to pick it up. Fourth, I think a cool voice change demo would be rad -- not sure what latency is like, but if you could get it end to end under like 250ms you could probably have someone exaggerate the mouth movements and pipe out audio of them talking like someone else. Which would be another really cool demo.
Super cool. You're going to be working 80 hour weeks for the next year, I'm excited to see what you put out!
The decoding is done via a companion app on your laptop or phone in the cloud, but end to end latency is still quite fast, under 500ms usually.
To begin with we don't plan to have a wake word, but rather a trigger key on your connected device that you're entering text into (pressing Alt for example), similar to Wispr Flow. We're doing some R&D into the silent speech equivalent of voice activity detection, so we have a low power basic RF detection always-on, not enough to decode speech from but enough to detect rough movement, and when it detects speech-like activity beginning to start we can trigger full sensing within tens of ms. This would likely ship as a software update a few months after shipping, though.
The voice change demo is something I've thought about actually, but we never got around to doing. Might film it soon :)
Have fun building this! It's worth doing.
Maybe I could have figured that out from the video but for folks who can’t watch the video it’s quite hard to discover what you’re building.
That's exactly where my mind went. Something I've wanted ever since I read that book a couple decades ago.
This is just really groundbreaking and I know the research has been out there for a little while but to actually implement is a whole nother ball game.
I think the use cases for this will expand as the product ships and gets out there!
"In the long, windowless hall, with its double row of cubicles and its endless rustle of papers and hum of voices murmuring into speakwrites, there were quite a dozen people whom Winston did not even know by name"
I find that I need this most often when I'm doing yard work, running equipment, or cleaning out the garage. Audio based microphones pick up so much noise.
There is a reason secure activity happens within secured rooms built to a specific standard https://en.wikipedia.org/wiki/Sensitive_compartmented_inform...
1. Insanely cool
2. Gives me serious "whispering earring" vibes. Imagine having a full on conversation with an AI via subvocalization and listening with ear bud to guide you throughout the day. https://gwern.net/doc/fiction/science-fiction/2012-10-03-yva...
I don't think it is an intuitive action at all and I became self-conscious about not making a sound rather than talking like I normally would.
I suppose it would require some "getting used to."
Regardless, I think it is a cool idea and a real problem. Sometimes I don't want to type nor talk out loud so I find myself whispering at my phone like I'm telling it a dirty secret.
--
As I'm writing this, I think this kind of product might not hit the market for consumers, but I'm seeing this used by Flock like companies to spy on people's speech, so kudos on making machine lip reading a thing!
1. https://github.com/mpc001/auto_avsr 2. https://github.com/amywork777/lipflow
It does feel a little weird to do it voluntarily when not emotional, but that'll be solved as soon as Claude fucks something up :)
I've found the opposite. Typing feels much more "deliberate", and of course you can hit backspace but you can't un-speak a word. With prompting, I've found that, in general, speaking leads to lower-quality results than typing.
> with many people
Not everyone. I, at least, definitely overthink what I type, and while I end up with nice-ish sentences in my prompts it takes a lot longer than a train of thought that the AI parses correctly most of the time anyway.
I suspect this has to do with the nature of the work you're doing; I dunno, just count me firmly in the camp of people who only rarely feel like I can express a thought sufficiently for coding agents with their voice.
But I guess thinking is going out of style anyway.
That said, I know a lot of people prefer talking to their computer, especially when their computer is a phone. So there is definitely a place for stuff like this.
Another trick is don't correct words and letters. Say the sentence you really meant to say. It is not uncommon for me to say a thought or a concept two, three, four times in a row as I'm wrestling with what the language should be on the written page.
But happy to see people get more options.
This reminds me of Eternity's End by Jeffrey Carver and the implants they had. Released in 2000 but the more time goes on the more the neural implant technology is coming about. To be fair I think Enders Game had similar tech to detect subvocal speech like this product.
Same. I'm often told I talk too fast, but I feel I communicate slowly because a lot of my fast speech involves halting and backtracking and such.
I can easily type 100WPM+, but I trivially blast out a few paragraphs of perfectly adequate text in a few seconds and then spend a few minutes re-editing it.
So this could take "noise canceling" to a whole new level if you could somehow "speak" across a loud bar or across the world, for that matter. Being integrated into something like Signal or similar end to end encrypted platform would be interesting as well.
Good reference! That series of books Is one of my all time favorites.
Second, I agree with exactly this. Something shaped like an AirPod
For about 15 minutes until they figure out what's going on, take those folks out back and "convince" them to spill the beans on how it work, then come up with a way to detect and prevent it moving forward.
https://github.com/joeblau/subvox
There was this really interesting one where it reads your lips:
https://github.com/amywork777/lipflow
I’m using these to build my own so I can work in co-working spaces, planes, and libraries. Though moving my mouth with no words might draw some eyes.
This is what I thought the original post was about. The glasses you have to wear is where that lost me.
If we could get lip-reading via your computer/phone camera down perfectly, that's the real win. Making people wear your goofy frames is where you lose a massive portion of people. Nobody wants to look like Zuckerberg lol
Lip reading AI would be pretty big since privacy would probably go out the window.
reminds me of tesla fsd v. waymo. may be a market for both? i.e., camera-only or camera + specific hardware.
Good luck on the project!
That's why we weren't able to make something that can mount to your laptop. We're trying to make a necklace form factor at some point in the future though, but can't give a date on this!
https://neuralink.com/updates/72-hour-voice-hackathon/
Edit: the more that I think about it the cooler and more important this feature seems!
Vowels change the shape of your mouth's cavity to change which frequencies to emphasize ("formants") and we hear the differences in vowels by analyzing those formant patterns.
I wonder if we could eventually end up with a "throat muscle" version where people learn to talk a new way without having to move the visible mouth muscles too.
The app essentially manages a bunch of TODO items written in markdown.
I think it's because most projects don't have any settled place (or at least, any place that lives inside the project's source-code repo) to record PRD-like information.
But because AI coding agents really do benefit from having PRD-like info (i.e. a hierarchical breakdown of the "intended what" and "intended how" of the system that "should" exist in the repo, kept updated to reflect changing ideas about those intentions), they really want to record that type of info somewhere, so they can then reference it.
AI coding agents don't seem to have any "taste" (or maybe "confidence"?) when it comes to asserting that even if you don't already have a separate PRD-like file in the repo, you should have one, if you want the agent to do its job well, and so it would request you to let it create one (or just automatically do it.) Instead, they just shrug and cram the PRD stuff into whatever structure you do already have. Which, if you're starting fresh with a greenfield project, is usually just a README.
(I hypothesize that if you tell an AI coding agent to use Behavior-Driven Development [BDD] on a greenfield project, it'll probably lean toward maintaining its PRD-like info in the form of BDD test cases. Haven't tried this experiment yet; would be interested to know the outcome!)
Throughout the document Claude kept claiming that issues were "mitigated but not fixed". They were essentially bandaid fixes for things it couldn't figure out. When I pointed this out to the author he said, "Well Claude thinks it works."
I wonder what happens to software when enough people try to ship projects like this.
I understood from just this. But the demo video impressed even further.
> Spent 30 seconds scrolling down to try to figure out what this is.
Did you try reading the site? Or maybe doing something other than just scrolling?
Does it need some form of subway surfer to get you to pay attention?
> Eventually clicked away from the site because it was taking too long to get to the point.
So yeah... subway surfer. This generation is cooked.
Throat mics don't work at all - see reviews online.
This is a different approach, so I am cautiously optimistic.
Edit: mostly concerned about RF near my eyes for 8 hours a day.
Also, the RF transmission only activates when you're actively dictating.
They get modulated since your mouth is a complex dielectric environment. Tissue, water, and air have different dielectric constants (different permittivity to RF) and the motion of the muscles in this medium leads to interactions with the waves that fill it, changing their phase and amplitude as they propagate.
Just preordered.
What languages do you support? (English is fine, but any others?)
Really like seeing the early prototypes here, and would love to see more on how the RF tech works on the website.
Are you able to patent it?
Regarding the website, you should lead with product, its details, rather than founders shown in 360 degrees from various vantage points. It is distracting and frankly irritating.
Calling it early - this is a scam. There's no way they're going to deliver that product, in that form factor, by Q1 2027.
And after some time I could not figure out what this is?
Put this part: " building smart glasses that let you talk to your computer without talking out loud. You can mouth words silently and discreetly and the glasses decode them to text at the speed of speech."
right at the top and maybe show someone actually doing that instead of hiding behind cryptic marketing.
Also, the product looks great.
Get your words out of the way 'say what it is'.
Very amazing tech though
Can we use RF for other medical applications? As an anesthesiologist, we do nerve blocks before surgeries or for pain. One rough indicator that it's working is muscle paralysis. Also similar for botox.
Very interested.
https://www.subtle.co
https://www.cnet.com/tech/services-and-software/apple-qai-ac...
Hope you guys are well-positioned patent-wise.
They use speckle patterns of a laser shone onto the skin, which is sensitive down to tens of nanometers. Pretty cool tech
From your comments it seems to reach this point you tried at least open source lip reading software, IMUs, variant antenna mounting strategies, various frequencies, and then gathered training data. You articulated strongly why the product is desirable and why the failed modes were dead-ends. I hope you can deliver on your pre-orders and are able to scale and/or sell the IP for a strong exit.
While this tech doesn't feel quite as dystopian as most recent AI innovations, I do think it is a further step in the direction of human disconnection.
Anyone who preorders - you are getting scammed.
Pangram: https://www.pangram.com/history/images/614ffe91-8ba9-4d3c-b1...
You can like or dislike generated images (if that's what they are—I haven't looked) but this post has nothing to do with scamming anyone, and given the wealth of technical information in the thread, including a video showing the product under development, that seems like a really weird claim to arrive at.
You don't see how this is misleading or fradulent (i.e. definition of a scam) in any way? A realistic image of someone wearing the glasses that don't exist yet, without disclosing it?
It's absolutely clear that if they're using AI generated images, they haven't even gotten a physical design of the product they're trying to sell.
That said, I think the idea of having a private walkie-talkie that connects on its own frequency for large crowded spaces to be very cool
"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."
https://news.ycombinator.com/newsguidelines.html
p.s. Especially please don't do this when a thread is brand new. Such comments are the easiest to post, and therefore the quickest to appear, but by posting them, you steer the discussion straight away from the curious, thoughtful conversation we want here. Threads are extremely sensitive to initial conditions, so this is important.
All of them are visionary and we would think that ppl gonna like them…
Well… ordinary people want boring tech.
This will be similar with whispering to glasses… Visionaries will like it… and return to missionary (boring) s*x (I hope that I don’t get a ban.)
Myself, I would buy these glasses immediately. But I would hate them as well as I would think I am insane (who wants to talk to yourself?).
Edit: Dang replied. I wish all the best to the team behind “whisper glasses”. It is amazing. Connect it with ElevenLabs and you rule the world. (Previous: It wasn’t meant a hate speech.)
A tasteless sex reference is one thing, but shallow-curmudgeonly dismissal of new work is the real sin here!
But I think (and see) that people prefer boring tech. Whisper glasses is amazing. But most people will probably find it creepy. (Pareto applied.). I hope I am wrong. These guys are like from sci-fi.
A comment like the GP quickly pattern-matches as snark and dismissal because those are so common on the internet. In other words, if you don't intend your comment that way, you need to do something to disambiguate this. Your follow-up comment does this nicely.
The reason this happens is that we all tend to take our intent for granted and forget that it exists in our head and isn't accessible to others unless we make it explicit. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
This also leads to a fascinating (to me at least!) phenomenon where the reply to a moderation comment (i.e. your follow-up here) makes explicit the thought which the commenter meant to communicate. What's fascinating is how different this is from what they originally posted. I call this the rebound effect. https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...