> Because the computational requirements of good speech rec exceed what you would build into a $100 device,

I have difficulty believing this. We had good speech recognition on <$1000 devices 10 years ago.

For unrestricted language, without training?

Yes to the former, pretty much for the latter - the "training" I recall was less than a minute long and consisted of repeating one or two example sentences, hardly anything that would be an adoption impediment to devices like an Echo. The resulting voice recognition worked, in my experience, better than that on modern Android phones.

I'm pretty skeptical of this. Apple has a pretty good commitment to privacy, one that is costing them on the AI front, and as far as I can tell they still upload Siri audio to the cloud for processing. If this was easy to do 10 years ago, I think Apple would at least offer as an option.

It's the reason everyone says Siri sucks - they do all the processing for her on the device, not the cloud.

Siri does more processing locally than Alexa does, but does upload to the cloud, and depends on the cloud, both for analyzing the statement and for accomplishing it. It can then execute local actions. The info uploaded to the cloud includes essentially all the relevant data for the query, so it doesn't particularly help with privacy -- it's just not the raw waveform.

But Siri sucks for MANY reasons, and not just technical ones. It does worse processing but also does worse cloud integration, and now vs. alexa, has a far worse ecosystem around it (alexa "skills" are pretty awesome, and trivial to create)

Pretty sure this isn't true. Siri won't work for me if I don't have an Internet connection.

That's not why people I know say Siri sucks: for us it's because it can't look up as much info as Amazon or Google. It understands the query, it just can't do as much with it.