A Listening Log for Generation Candidates
Log candidate files, timestamps and actionable revisions rather than selecting by first impression.
Screen requirements before preferences
First reject missing sections, incompatible duration, unwanted vocals or unusable exports. Compare the remaining files at similar perceived playback loudness. Louder should not automatically score better. Preserve the originals and use listening gain for comparison rather than overwriting source levels.
Reusable candidate log
| Candidate | Tool, version and actual settings | Problem and time range | Decision and next action |
|---|---|---|---|
| A: local file | Prompt, date; seed only if supported | Narration is masked at 00:12–00:15 | Keep the structure, reduce density |
| B: local file | Another candidate for the same brief | Abrupt ending at 00:28 | Extend, edit the tail or replace |
| C: local file | Attempts and actual cost | Good melody, but missing required stems | Resolve the delivery constraint |
These are fictional teaching rows, not measurements of a model. Assess form, performance or vocals, audible artefacts, editability and deliverables separately. Pair each criticism with a timestamp and a feasible action. “Not professional enough” does not define an edit.
Recheck and stop deliberately
Listen to the whole piece, then loop problem sections. Revisit after a break or involve another listener. Changing one main requirement helps record keeping, but a single successful candidate may still be random variation. Do not treat it as proof that a phrase always works. Proceed to editing when requirements are met and remaining repairs fit the budget. Regenerate when essential structure fails and local revision is unavailable. Set candidate limits from the project's budget and deadline.