Put In a Good Word

The sound of words is always an issue in songs because they can be awkward for singers. For some listeners, some words are fingernails on a blackboard. For AI "singers," they might never be able to pronounce them, or they hallucinate them, which can sometimes be interesting, or they will put emphasis on the wrong syllables, like "geometry", which AI would pronounce as "GE-om-e-try" with a southern lilt, or ge-om-e-TRY, instead of ge-OM-e-try. 

"Quaternity" was also a word it could not pronounce, perhaps because there are no word patterns (embeddings) that start with "quat". It's usually "quo" or "quar" or "quin". 
In some cases, the sound of the word itself doesn't matter, particularly in protest songs--like Springsteen's Streets of Minneapolis. The city could be any city, like "Park Forest," and it doesn't matter how musical it is. "Chicago" is famous for its word, an "upbeat" word falling on the 4th beat of a bar or the "and" of 4.

Displacing syllables can be both awkward and interesting. You can use a word as a placeholder for its rhythmicity, then swap the words out, or just use it as a rhythmic figure, which is my idea of the wordrum, a word used for its rhythms as "drumming" instead of what it means.

 On embeddings and TTS
 
The Music of Language "Discuss speech rhythms"