Have fun building this! It's worth doing.
If you were actually serious about prioritizing privacy above securing an ongoing service relationship you would ship local inference as the default and only (possibly later) offer a cloud service as an optional add-on for those who want/need it.
You can get both markets (ease of use and technical) by allowing it. Also i would bet that you gain almost no product security locking off local inference - the hardware is the hard bit. Once someone has your hardware in hand it will not take them long to reverse and build their own processing layer; you’ll need to fight those people on patents and product quality.
Heck for a v1 just allowing local sensor data would probably be enough for the tech enthusiasts. But please don’t lock this off from the get go - it makes the sale harder and I don’t think it gives you much in return.
Cloud processing can be tempting as the primary revenue stream, but it entirely cuts off those users who cannot send arbitrary PII over the internet (regardless of encryption promises).
It also invites the (especially now-LLM-empowered) enthusiasts to reverse engineer it.
Instead, consider selling the glasses with >20% profit, and give local users a cheaper subscription to something else (maybe the model updates?) so they can also help keep the company sustainable.
Why don't you do this on the phone? You don't want to have your proprietary tech on the user device? Or maybe you wanted to save the recording on your cloud anyway so you can train on it later or otherwise use the data?
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 :)