The presentation of this product reminds me of peak crypto when a 'white paper' and a two-page website was all anyone needed to get bamboozled into handing their money over.
The accuracy is going to be the real make or break for this. In a paper from 2018 they reported 92% word accuracy [1]. That's a lifetime ago for ML but they were also using five facial electrodes where now it looks confined to around the ears. If the accuracy was great today they would report it. In actual use I can see even 99% being pretty annoying and 95% being almost unusable (for people who can speak normally).
I found it interesting that in the segment where two people were communicating "telepathically", they seem to be producing text, which is then put through text-to-speech (using what appeared to be a voice trained on their own -- nice touch).
I have to wonder, if they have enough signal to produce what essentially looks like speech-to-text (without the speech), wouldn't it be possible to use the exact same signal to directly produce the synthesized speech? It could also lower latency further by not needing extra surrounding context for the text to be pronounced correctly.
The great thing about a product like this is that it's so easy to fake it in video.
I don't really buy that typing speed is a bottleneck for most people. We can't actually think all that fast. And I suspect AI is doing a lot of filling in the gaps here.
It might have some niche use cases, like being able to use your phone while cycling.
I agree that it's an easy to fake demo and at the same time, if they're going to fake it, why make it seem so slow?
As to whether typing speed is a bottleneck for most people, maybe not most people, but definitely some people, and it's a massive bottleneck for me personally.
I think better when I'm talking and since I have started using speech to text, it has increased my writing speed and coding speed by at least an order, maybe two orders of magnitude.
But you are right, the AI filling in gaps can really cause trouble using speech, goodness knows what it's doing using sub-speech.
>I agree that it's an easy to fake demo and at the same time, if they're going to fake it, why make it seem so slow?
Honestly I have no idea if it's fake. I wouldn't be surprised if it's both fake and real: the actual video is entirely fake, but a reasonably accurate demonstration of actual capabilities (like a lot of tech demos at live events...)
One of the major ways you can speed up reading, is that you stop 'vocalizing' each word in your head. It does seem that thinking is much faster than 'thinking aloud' (in your head)
So this a bone conducting microphone? That operates at the speed of speech? While you sit around awkwardly, hoping no one talks to you? This isn't thought. This is you saying to yourself quite clearly what you would like it to hear.
I just imagine this going really wrong. My chain of thought would be something like: "Let's see, I need to rotate this image so I need to loop over rows then columns, .. gawd fuck this code base is shit designed, there are no units on these fields, this could be so much cleaner, ... for each row ... I wonder what's for lunch today? I hope it's good ... for each column ... Dang that response on HN really pissed me off, I'd better go check it ... read pixel from source ... tonight I meeting up with a friend, I'd better remember to confirm, ... write pixel to dest ...."
What I picked up from this vision of the future... we will have mind reading devices to capture out thoughts, but we will still be on a train and commuting to work... dang...
So they came up with this groundbreaking idea but couldn't come up with better use case then typing on a train.
Look, I can't but not appreciate that at least they are doing something interesting as opposed to vibe one shot fork of vs code things that we see.
> We currently have a working prototype that, after training with user-specific example data, demonstrates over 90% accuracy on an application-specific vocabulary. The system is currently user-dependent and requires individual training. We are currently on working on iterations that would not require any personalization.
"AlterEgo reads information from the peripheral somatic system through internal speech movements, rather than directly from the brain. It detects the signals users send to their mouth and vocal cords when deliberately, but silently, voicing words. "
This is literally only as fast as text to speech. the only difference is that you don't have to speak aloud. Which is cool.
But for using a computer its still annoying and worse than a mouse because with a mouse you can click or drag and place in a second, in this format you have to think "move the box from point A to point B (with coordinates or a description) etc etc".
I think its cool, I've been brainstorming how a good MCI would work for a while and didn't think of this. I think its a great novel approach that will probably be expanded on soon.
For those thinking about speed: an average human talks anywhere from 120-240 words per minute. An average human who touch types is probably 1/3 to 1/2 as fast as that, while an average human on a phone probably types 1/5 as fast as that.
But for me speed isn't even the issue. I can dictate to Siri at near-regular-speech speeds -- and then spend another 200% of the time that took to fix what it got wrong. I have reasonable diction and enunciation, and speech to text is just that bad while walking down the street. If this is as accurate as they're showing, it would be worth it just for the accuracy.
As someone with ADD and a lot of crosstalk in my "inner voice", I can't imagine this could make any sense of what I intending, let alone one thing. Definitely a lot of use cases if it isn't vaporware.
56 comments
[ 4.2 ms ] story [ 77.5 ms ] threadhttps://www.media.mit.edu/projects/alterego/overview/
adding also their press release here:
https://docsend.com/view/dmda8mqzhcvqrkrk/d/fjr4nnmzf9jnjzgw
I suspect it's EMG though muscles in the ear and jaw bone, but that seems too rudimentary.
The TED talk describes a system which includes sensors on the chin across the jaw bone, but the demo obviously has removed that sensor.
[1] https://www.media.mit.edu/publications/alterego-IUI/
I have to wonder, if they have enough signal to produce what essentially looks like speech-to-text (without the speech), wouldn't it be possible to use the exact same signal to directly produce the synthesized speech? It could also lower latency further by not needing extra surrounding context for the text to be pronounced correctly.
I don't really buy that typing speed is a bottleneck for most people. We can't actually think all that fast. And I suspect AI is doing a lot of filling in the gaps here.
It might have some niche use cases, like being able to use your phone while cycling.
also adding their press release here:
https://docsend.com/view/dmda8mqzhcvqrkrk/d/fjr4nnmzf9jnjzgw
depends on what they are connected to in the back there.
As to whether typing speed is a bottleneck for most people, maybe not most people, but definitely some people, and it's a massive bottleneck for me personally.
I think better when I'm talking and since I have started using speech to text, it has increased my writing speed and coding speed by at least an order, maybe two orders of magnitude.
But you are right, the AI filling in gaps can really cause trouble using speech, goodness knows what it's doing using sub-speech.
Honestly I have no idea if it's fake. I wouldn't be surprised if it's both fake and real: the actual video is entirely fake, but a reasonably accurate demonstration of actual capabilities (like a lot of tech demos at live events...)
One of the major ways you can speed up reading, is that you stop 'vocalizing' each word in your head. It does seem that thinking is much faster than 'thinking aloud' (in your head)
So they came up with this groundbreaking idea but couldn't come up with better use case then typing on a train.
Look, I can't but not appreciate that at least they are doing something interesting as opposed to vibe one shot fork of vs code things that we see.
https://www.media.mit.edu/projects/alterego/frequently-asked...
Seems like vaporware.
I think its cool, I've been brainstorming how a good MCI would work for a while and didn't think of this. I think its a great novel approach that will probably be expanded on soon.
But for me speed isn't even the issue. I can dictate to Siri at near-regular-speech speeds -- and then spend another 200% of the time that took to fix what it got wrong. I have reasonable diction and enunciation, and speech to text is just that bad while walking down the street. If this is as accurate as they're showing, it would be worth it just for the accuracy.
Going from voice input to silent voice input is a huge step forward for UX.