@AndrewFG
Hmm. Just my personal experience is that the recognition isn't "appallingly bad", though it isn't perfect, and is inconsistent. But generally, I can get what I want with not trouble. (For the most part, for me, the difficulties have to do with foreign names (so for classical music, yeah, it's not good).
But, obviously, your mile might (and does) vary.
However, simply adding the metadata set from your library likely won't help. The reason is that once the utterance samples get numerous enough, adding more does not help the recognition. I can't remember how many utterance samples I had for the early versions of HouseBand - but thousands. Adding more didn't help. The best recognition comes when there are few samples, and the user speaks on of them.
More and more Amazon is moving away from the model where the developer provides samples towards a model where the developer just lets Amazon know what *kind* of utterance it is (e.g., mentioning a pop artist, or the name of a US city), and then handles more and more on their end.
But I think the accuracy will continue to improve.
What would help more is not if the current tags could get uploaded to Amazon (because adding a tag like "10 d E A T h b R E a s T ⚄ ⚄" -- a track from the recent Bon Iver album -- won't help anything.) What would be better is if music started including voice accessible tags.
I was thinking of trying to write a MC script convert tags to their metaphone equivalents automatically, create new fields for those tags, and then have HouseBand search those tags. I think this would make a huge improvement.
Steve
Steve