We just need someone insightful and ingenious to find a way to deal with machine learning in an 'offline' way, and be able to present the user interface in a quick fashion.
It would have to start out very dumb, but with some great key algorithms I expect an open source option could move a lot faster than anything out there in this regard.
Precisely. I don't get what the misunderstanding is here among the Slashdot crowd.
Natural Language Processing is neat tech. Mechanics of speech recognition is neat tech. Integration of the two via a dispatch engine and scriptlets to go off an search Google, run a command, or whatever else one can script, is neat tech.
I'd use this ALL THE TIME if the data didn't leave my network, and I'm sure I'm not alone.
We can't duplicate a zillion far off machines running a Google-scale cluster, but it's hard to see why we need to in OSS land. I have a spare 32 core box and God-knows-how-many GPUs sitting here. Where's the project that can let me get up and running on my own, and that we can all use to iterate over as public algorithms (inevitably) improve and storage/memory/processing costs (inevitably) decrease.
Frankly, it's difficult to see why that type of infrastructure is really needed in the first place. NLP is hard, but it's not like these building blocks aren't already there. Apple's dictation software (PlainTalk) was running on System 7.1 Pro 20 years ago, using local hardware 100's of times slower than what I have in my pocket. Basic NLP code was running on the Newton, which was 1000x slower and still managed to handle the basics on top of the handwriting recognition. "Speakable Items" let me run user-writable AppleScripts to automate tasks and was just missing dictatable variable names.
None of this required cloud-level processing, especially not voice recognition, which even Apple lets you do locally w/o using Siri.