Mv Jp Lyrics Subtitle — Development skill for Claude Code

mv-jp-lyrics-subtitle

An agent skill for putting **Japanese song lyrics on screen the way a Japanese reader reads them** — not the way a renderer happens to wrap text.

Four mistakes this fixes, all of which "look fine" to a non-Japanese viewer and read as wrong instantly to a Japanese one:

Mistake	What it produces	Fix
Timing by syllable count	every mora gets the same time; real singing is 2–3× uneven	measure onsets on the separated vocal track
Breaking mid-word	君 / を見るたび — a particle alone at line start	break at 文節 (bunsetsu) boundaries only
All-kana lyrics	the AI-generated lyric sheet written in hiragana, pasted straight to screen	漢字仮名交じり normalisation
Snap-to-pitch-change	the phrase-final falling note is read as the next phrase's start, ~0.4 s early	energy gate: singing must continue for 180 ms

Install

Drop the folder into your agent's skills directory, or add it as a submodule. No build step, no dependencies beyond `librosa` / `numpy` for the measurement script.

Use

# 1. measure real onsets from a separated vocal track
python scripts/measure_lyric_onsets.py --vocal vocals.wav --lines lines.json --out timing.json

# 2. gate the render on the two content checks
python scripts/check_lyric_coverage.py --lines lines.json

`lines.json`:

[{"id": 9, "start": 42.26, "end": 48.00,
  "kana": ["きみを", "みるたび", "しんぞうがとまるの"],
  "disp": ["君を", "見るたび", "心臓が止まるの"]}]

Output flags any line whose snap deviation exceeds `--unsure 0.45` seconds, so you know which lines still need a human listen instead of shipping silent errors.

What is in SKILL.md

• the eight hard rules (R1–R8) with counter-examples
• the three-stage onset algorithm and why stage 3 cannot be skipped, with the measured false positive (44.81 accepted vs 45.24 true)
• a kana → kanji normalisation table, and where not to over-kanji-ise
• typography numbers reverse-engineered from four commercial anime MVs (97–108 px