Yes, typical mics tend to pick up the typical human frequency range, though cheaper mics may have some really poor characteristics at the edges. Usually in the speech range they'll be pretty solid.
However, there's a lot of play within the space. One difference is that microphones do a very direct recording of the sound waves, but what we hear is actually very distorted compared to the "real" sound by the nature of our ear. One of the big differences is that if there is a very loud 4000Hz sound, we can't hear a soft 4005Hz sound near it very well, but the microphone "hears" it just fine. So for instance, you could put out a loud sound for a user, but embed a very quiet command in frequencies the human couldn't hear, but if the listening model doesn't account for that (and there are reasons it wouldn't necessarily want to, because it wants to hear commands even in the presence of significant background noise), you could get commands in to a system. See https://en.wikipedia.org/wiki/Psychoacoustics for discussion about how our ears fail to pick up the "real audio" signal, and how much we've exploited that in music compression.
Now, that was a very brute force example. It sounds to me like what this article is talking about are called "adversarial examples" (https://blog.acolyer.org/2017/02/28/when-dnns-go-wrong-adver... ). Voice recognition doesn't listen the same way we do, it doesn't necessarily take a holistic view of the signal, but is looking for specific frequency patterns and changes and turning that into phonemes, into words, etc. (There's a lot of ways of doing this and I don't specifically know what Alexa and Siri are doing, so that's a really vague overview.) If you know what they are looking for, you can use filters to very, very selectively remove the patterns from a bit of music or something that Alexa might trigger on, and then insert just the bare minimum skeleton of the sounds that it is really recognizing. A human won't be able to hear the difference (most likely; depends on how badly the original is mangled but even if it is audible it is almost certainly not audible without an A/B test and very good ears), but the probably-neural-nets monitoring for sounds will end up superstimulated and interpret the adversarial example as words.
While the adversarial examples work best with tuning to the target network, widely-shared networks like Alexa or Siri mean that such tuning is practical where attacking some custom-trained model used by one person isn't, and experiments have shown that adversarial examples travel between separately-trained nets and even non-neural-net models to a much, much greater degree than what at least my own intuition would have suggested before hand. (See previous link and look for the discussion of "Practical black-box attacks against deep learning systems using adversarial examples". It is extremely counter-intuitive to me how easy this is.)