Honestly, I’m disappointed with the quality of the generated tracks so far. I spent 30 credits and didn’t get a single acceptable result.

In about 80% of generations the first word sounds garbled, as if something’s stuck in the mouth; from the second word onward it’s fine. Stress and pronunciation are problematic — the system often places stress incorrectly, even on month names. Sometimes entire verses are missing and the song starts with the chorus.

It’s unclear which language model is being used — it feels like models are selected randomly, so some tracks have a noticeable accent.

I’ll try tweaking the settings more, although I feel I’ve already tried everything available.