notes: judge low confidence against the song, and flag unsingable words #28

2026-10-10T23:33:23-07:00

dejima-agent commented

2026-10-10 23:33:23 -07:00

Summary generated by LLM agent

Fixes #17.

On sung vocals the low-confidence flag marked 73–99% of notes, so neither the review's highlighting nor the end-of-run warnings pointed anywhere in particular.

What the data showed

Measured on two real songs' existing alignment and pitch files, without running any model. As known-bad words I took ones nobody could sing as aligned: under 0.08s, stretched over 3s, or under 25% voiced.

The alignment score does separate bad words from good ones. It ranks a bad word below a good one about 78% of the time on both songs, and bad words' median score is 6–10× lower.
The score's scale differs a lot between songs: medians of 0.08 and 0.28. A fixed cutoff can't work for both.
The score alone catches only about half the known-bad words. The rest are confidently aligned but impossible, such as a word stretched across an instrumental break.

Rule

A word is low-confidence if:

it scores below 0.25× the song's median score, or
it lasts under 0.08s or over 3s.

Words with no voiced pitch are flagged as before. Flagged notes less than 2s apart now share one warning.

	notes flagged	low-confidence warnings
song A, before	99%	a few runs spanning most of the song
song A, after	33%	21
song B, before	73%
song B, after	35%	29

Song A's 10-second stretched word (the subject of #15) now gets its own warning, "20 low-confidence notes near 1:22–1:32", where before it was lost in a song-wide run.

Testing

test_build_notes_low_score_word_flagged is replaced. It flagged a lone word for scoring 0.1, but under a relative rule a single word is its own median.
New tests: a word scoring far below the rest is flagged; a uniformly low-scoring song isn't; words too short or too long to sing are flagged whatever their score; flagged notes close together share one warning.
Full suite passes, except test_version (needs the package installed).

Existing songs pick this up with Restart from set to notes, or for every song at once with REBUILD ALL FROM… notes.

**Summary generated by LLM agent** Fixes #17. On sung vocals the low-confidence flag marked 73–99% of notes, so neither the review's highlighting nor the end-of-run warnings pointed anywhere in particular. ## What the data showed Measured on two real songs' existing alignment and pitch files, without running any model. As known-bad words I took ones nobody could sing as aligned: under 0.08s, stretched over 3s, or under 25% voiced. - The alignment score does separate bad words from good ones. It ranks a bad word below a good one about 78% of the time on both songs, and bad words' median score is 6–10× lower. - The score's scale differs a lot between songs: medians of 0.08 and 0.28. A fixed cutoff can't work for both. - The score alone catches only about half the known-bad words. The rest are confidently aligned but impossible, such as a word stretched across an instrumental break. ## Rule A word is low-confidence if: - it scores below 0.25× the song's median score, or - it lasts under 0.08s or over 3s. Words with no voiced pitch are flagged as before. Flagged notes less than 2s apart now share one warning. | | notes flagged | low-confidence warnings | |---|---|---| | song A, before | 99% | a few runs spanning most of the song | | song A, after | 33% | 21 | | song B, before | 73% | | | song B, after | 35% | 29 | Song A's 10-second stretched word (the subject of #15) now gets its own warning, "20 low-confidence notes near 1:22–1:32", where before it was lost in a song-wide run. ## Testing - `test_build_notes_low_score_word_flagged` is replaced. It flagged a lone word for scoring 0.1, but under a relative rule a single word is its own median. - New tests: a word scoring far below the rest is flagged; a uniformly low-scoring song isn't; words too short or too long to sing are flagged whatever their score; flagged notes close together share one warning. - Full suite passes, except `test_version` (needs the package installed). Existing songs pick this up with **Restart from** set to `notes`, or for every song at once with REBUILD ALL FROM… `notes`.

dejima-agent added 1 commit 2026-10-10 23:33:24 -07:00

notes: judge low confidence against the song, and flag unsingable words

tests / test (pull_request_target) Successful in 4m18s

Details

c0b782baa9

A fixed 0.5 alignment-score cutoff flagged 73-99% of notes on sung
vocals, so the flag pointed nowhere. Scores still separate bad words from
good ones, but their scale differs from song to song: flag a word below a
quarter of the song's median score, or one shorter than 0.08s or longer
than 3s, which no singer produces and which a word stretched over
something else does. Flagged notes less than 2s apart share one warning.

branderson approved these changes 2026-10-11 00:45:39 -07:00

branderson merged commit 8739033cf7 into main

2026-10-11 00:45:45 -07:00

branderson referenced this issue from a commit

2026-10-15 14:17:08 -07:00

Merge pull request 'notes: judge low confidence against the song, and flag unsingable words' (#28) from relative-confidence into main

dejima-agent referenced this issue from a commit

2026-10-15 17:44:37 -07:00

Merge pull request 'notes: judge low confidence against the song, and flag unsingable words' (#28) from relative-confidence into main

Sign in to join this conversation.