4 ms·

That's because it's using the phones processors, no? In a hypothetical fbi is recording scenario they might have more sophisticated language processing

• tinus_hn 3y ago
  No. If you implement these systems, the model outputs the sentence it guessed and the confidence the user said that sentence. The activation logic is tuned so it triggers on pretty low confidence so the more sophisticated system can try working on the actual command.
• godelski 3y ago
  No, it is because the problem is much easier. Voice recognition doesn't just mean amplifying the sounds (there's only so much of that you can reasonably do), but rather discriminating a phrase from other phrases. You only have two classes: "hey siri" and everything else. This is much easier than differentiating a conversation, where words and phrases are not so easily distinguishable: "Sam went and got Tom" vs "Man went and got bomb". To put this in an easier context, let's say we have colored shapes of various sizes. We place them down and start walking away. It's probably easier to discern color than shape at a distance (is that a hexagon or octagon?). And sufficiently far away it is hard to distinguish colors that are near one another (is that brown, firebrick, or maroon?). You can probably see that the complexity of the task is critical and that discrimination is easier when you have fewer bins to place things in.
• DANmode 3y ago
  There is a separate chip for just this, but yes.