Earlier quoted context omitted.
Because different words have very different best potential performance characteristics when used as wake words. We would have liked to enable fully custom wake words, but it would have led to bad user experiences with the device either failing to wake up too often (false negatives) or having too many false positives or both. So we compromised by enabling more than one, but not full choice.
Thanks for the insight. Is it difficult to fingerprint a sound bite? Have people repeat it several times and again for each user that wants to use the custom wake word?
Type 2 words make better potential wake words.
(That's not the whole story, but I think it's good model.)