I built an off-line voice recognition system 10 years ago using Microsoft Speech API and I guarantee you it worked quite well. There's nothing in base voice recognition that would require uploading everything to the cloud. You only have to be willing to read out some text for few minutes, once, to train the model.

> You only have to be willing to read out some text for few minutes, once, to train the model.

I guess that's what kills it as practical in the real world.

Laziness is competitive advantage. There are costs to it though - in this case, we pay for it with our data. Small bits of our souls, if you will.

I don't expect these technologies to appeal to people who think of the data their behavior generates as "small bits of their souls."

For everyone else though, there are some neat families of features in all the major tech stacks.

"We need real-world audio samples" does not entail "let's collect audio [and behavioral] data from one billion users". There are less intrusive ways to collect data. How about paying specific individuals for their data. For example, there are about 5000 Nielsen families, http://entertainment.howstuffworks.com/question433.htm.

For a company that supposedly is employing the best minds in statistics, the inability to effectively use statistical samples is, uhm, intriguing.

With the inception of digital cable boxes, many cable providers no longer need the Nielsen data because the cable box tells the provider what channel it's tuned to (a technical necessity if the box is using switched digital video). It's much easier (and safer for the end user) to passively collect usage data via the application you provide than to trust some third-party to (a) collect a statistical sample unbiasedly and (b) collect the sample securely while protecting user privacy.

Even if Google is using statistical sampling, it's still something they'd want to collect directly through the app, not via a third-party.

As for paying individuals for their data, Google does that too. https://play.google.com/store/apps/details?id=com.google.and.... But if your argument is "They should be paying everyone they collect passive samples from during use of their products," that's an argument over price-point, not over whether the collection itself is right.

The rightness question seems to me to look a lot more like "If you can collect data from a billion users without doing any harm to the users, and that data is going to be more useful than a statistical subset of that data, why should you not collect it?"

> The rightness question seems to me to look a lot more like "If you can collect data from a billion users without doing any harm to the users, and that data is going to be more useful than a statistical subset of that data, why should you not collect it?"

In this particular case (speech recognition), one harm done to the users is tying the product to the Internet, and thus requiring what should be closed-loop tasks to go through vendor's servers.

"Should be closed-loop" is an interesting assumption that I'm not convinced aligns with the reality of the technology. To what extent has speech recognition been improved by being able to feed it through a constantly-updated set of open-loop ML infrastructure?

Probably a lot; open-ended voice recognition with no prior training is hard (though I'm not convinced it can't be made to work off-line now that the models are there). Still, a lot of devices don't need open-ended voice recognition (structured grammars greatly simplify the problem), and if you allow for users having to pre-train their devices for few minutes, off-line processing becomes easier.

My impression is that the main driver behind cloud-first voice processing is business, not technology.