aarushyadav

8-15% on an unseen user as of now. We are scaling our pretraining massively before we ship out the units and will likely be 3-5% by then, which is comparable to most ASR models in real-world conditions.

Not immediately, no. There are also different regulatory processes to navigate in Europe for RF approval, so I can't give you a timeline yet unfortunately.

- No, from the beginning we will have a free tier with a cap on the amount of words you can dictate, similar to Wispr Flow, but with generous limits for most users. And a paid tier for ~$20/mo with unlimited dictation - $20/mo for unlimited dictation - Yes, a few months after shipping the glasses we plan to ship a small local model - Correct

You can reach out at [email protected] if you have any other questions, happy to answer :)

Amazing! Supports just English for now, but we will keep pushing out other languages as software updates after shipping, with Spanish as the second planned language.

That's fair, will rethink!

Tbh we haven't tried internal speech on outsiders yet, just the two of us and one of our employees. We were able to solve it on that small set, so it's technically solvable with enough user specific internal speech personalization data. But the amount of time it took us to learn the behavior comfortably is a bit too high for us to focus on it as a key feature we have from the beginning. In a few months though we will start collecting larger scale data on internal speech as well.

Additionally, we found that you can mouth words discreetly enough / with no exaggerated movements, such that people sitting around you would likely not notice unless they were staring at you, which seems to be good enough, as a start at least.

We did consider that, but it's a bit hard to fit all the electronics well and distribute the antennas optimally and keep that consistent across people and glasses. Best to have a predictable device we can engineer to the needs of this application!

Yep. You can either just dictate silently and we can transcribe, or another feature we will likely push out as a software update is where you can speak out loud and the RF signal is used to fully and almost perfectly denoise the audio of you speaking.

Fair enough! We take data security and privacy very seriously and have built many safeguards to prevent even us from accessing the raw data in the cloud. At some point in the future we hope to launch local inference models too!

Glasses are a great way to provide output back, via an AR display, and also offer more real estate to fit the electronics and antennas in.

Yes, this is the unoptimized speed while we are in development. We are working hard to make it much faster for production!

When we launch, no, it will be cloud inference, but will be processed with encryption in hardware isolation and only retained and used for anonymized training with your consent. At some point in the future we may make models available for local inference though.

English only right now, plan to add support for others after shipping

this should be fixed now!

Not well enough! You can get to 15-20%ish WER but going beyond that doesn't seem feasible with the info available from a camera, even with a great language model