Builders often collapse “finish the song” into one afternoon: rewrite the chorus, slam a limiter, export. Then the hook changes again and the mastered file is trash. This how-to is a short sequencing guide—lyric lock first, mastering last—so your pipeline stays reviewable. A suite like WAV AI is useful when those steps stay named, not when they blur into one button.
Pipeline contract (write this once)
- Language lane — verses / chorus / bridge text only.
- Music lane — generate or edit bed against locked text.
- Delivery lane — mastering / loudness / width for the export.
Never open lane 3 while lane 1 is still contested in Slack.
How-to: Lyric Studio pass
Use Lyric Studio as the language lane:
- Paste the song idea + section map (V1 / Pre / Chorus).
- Generate two drafts from different model choices if available; keep diffs explicit.
- Revise only the lines that fail the brief (POV, product claim, syllable count).
- Read the keeper aloud against picture or VO scratch—silent reading lies.
- Freeze a
lyrics-vN.md(or equivalent) before any mastering click.
Text-to-music research such as MusicLM (Google Research) shows language can steer hierarchical generation. That is a reason to lock language before you spend mastering listens—not after.
How-to: mastering pass (only after freeze)
Open AI Music Mastering when lyrics + bed are approved:
- Import the keeper bed (not yesterday’s lyric experiment).
- Preview under dialogue at target platform loudness habits; watch for pumped breaths.
- Prefer one mastering candidate that keeps speech intelligible over the “loudest” export.
- Name the file with date +
mastered; archive the pre-master.
Loudness practice for broadcast/streaming is documented in EBU R 128. Treat AI mastering as a fast path to that discipline, not a license to ignore dialogue.
Anti-patterns
- Mastering three lyric variants “just in case.”
- Reopening Lyric Studio after you already limited the bed hard.
- Citing product marketing stats as “data”—use third-party references like Magenta’s public research culture at magenta.withgoogle.com for tooling context instead.
Ship when speech reads clean and the lyric log matches the audio. Polish theater is optional.

Top comments (0)