LOW_SCORE = 0.5 in notes.py marks a note low-confidence when its word's alignment score is below 0.5. On sung vocals almost every word falls below that line, so the flag no longer singles out the notes that need checking.
On one real song, 99% of the 450 aligned words scored below 0.5. The scores ran 0.00 / 0.08 / 0.31 at the 10th / 50th / 90th percentile, and 541 of 542 notes were flagged. The score is the MMS_FA aligner's mean per-frame probability, and that sits far lower for singing over a separated stem than for the clean speech the 0.5 was presumably chosen for.
Options:
set the threshold relative to the song, for example its bottom 10–20% of scores
calibrate a fixed value from several songs
combine the score with other evidence, such as how much of a word's time is voiced
Either way the review and the "N low-confidence notes near …" warnings should go back to pointing at the worst spots.
**Filed by LLM agent**
`LOW_SCORE = 0.5` in `notes.py` marks a note low-confidence when its word's alignment score is below 0.5. On sung vocals almost every word falls below that line, so the flag no longer singles out the notes that need checking.
On one real song, 99% of the 450 aligned words scored below 0.5. The scores ran 0.00 / 0.08 / 0.31 at the 10th / 50th / 90th percentile, and 541 of 542 notes were flagged. The score is the MMS_FA aligner's mean per-frame probability, and that sits far lower for singing over a separated stem than for the clean speech the 0.5 was presumably chosen for.
Options:
- set the threshold relative to the song, for example its bottom 10–20% of scores
- calibrate a fixed value from several songs
- combine the score with other evidence, such as how much of a word's time is voiced
Either way the review and the "N low-confidence notes near …" warnings should go back to pointing at the worst spots.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Filed by LLM agent
LOW_SCORE = 0.5innotes.pymarks a note low-confidence when its word's alignment score is below 0.5. On sung vocals almost every word falls below that line, so the flag no longer singles out the notes that need checking.On one real song, 99% of the 450 aligned words scored below 0.5. The scores ran 0.00 / 0.08 / 0.31 at the 10th / 50th / 90th percentile, and 541 of 542 notes were flagged. The score is the MMS_FA aligner's mean per-frame probability, and that sits far lower for singing over a separated stem than for the clean speech the 0.5 was presumably chosen for.
Options:
Either way the review and the "N low-confidence notes near …" warnings should go back to pointing at the worst spots.