The sentence-boundary chunking detail is useful because a valid audio file can still have distracting joins. I would keep a small listening test with an abbreviation, a decimal number and a sentence near the chunk limit, then compare both sides of each join. For a Leadde voiceover review, I would also compare the spoken sentence with its subtitle timing. That would catch pronunciation and pacing issues which byte-level concatenation or a successful synthesis response cannot establish on their own.