got-feedBack/feedBack
1
0
Fork 0
mirror of https://github.com/got-feedBack/feedBack.git synced 2026-08-10 18:59:56 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix(gp_autosync): slope-constrained DTW steps — stop path collapse on riff-based songs (#791)

Browse Source 
librosa.sequence.dtw's default step sizes permit pure horizontal/vertical
moves; on songs whose chroma is self-similar for long stretches the flat
cost surface let the path collapse (minutes of score onto one audio frame),
so auto-sync produced monotonic-but-garbage sync points and the per-bar
warp imported charts badly out of sync while reporting success.

Use the standard music-sync step pattern [[1,1],[1,2],[2,1]] (local tempo
ratio bounded to 0.5x-2x), falling back to unconstrained steps if the
global length ratio makes it infeasible.

Validated on the reported song (138 BPM tab, YouTube audio): coarse points
now track 1:1, refine holds slopes 0.77-1.04, warped downbeats hit onset
peaks at 3.3x background energy.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Byron Gamatos
2026-07-05 23:12:31 +02:00
committed by GitHub
co-authored by Claude Fable 5
parent d567fd5597
commit 1a8540935b
2 changed files with 27 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
lib/gp_autosync.py
+24 -1
View File
@@ -538,9 +538,32 @@ def _dtw_align(
Returns wp where wp[i] = [score_frame_index, audio_frame_index].
"""
import librosa
import numpy as np
cs = _safe_normalise(chroma_score)
ca = _safe_normalise(chroma_audio)
_D, wp = librosa.sequence.dtw(cs, ca, metric='cosine')
# Slope-constrained step pattern ([[1,1],[1,2],[2,1]], Müller's standard
# music-sync config): every step advances BOTH axes, bounding the local
# tempo ratio to 0.5x-2x. librosa's default steps allow pure
# horizontal/vertical runs, and on riff-based music (long self-similar
# chroma stretches, e.g. stoner/doom) the flat cost surface let the path
# collapse — whole minutes of score mapped onto a single audio frame,
# producing garbage sync points. The constrained pattern makes that
# degenerate path impossible.
steps = np.array([[1, 1], [1, 2], [2, 1]])
weights = np.array([1.0, 1.0, 1.0])
try:
_D, wp = librosa.sequence.dtw(
cs, ca, metric='cosine',
step_sizes_sigma=steps, weights_mul=weights,
)
except Exception as exc:
# The constrained pattern needs the global length ratio within its
# 0.5x-2x slope bounds; a pathological pairing (e.g. a 3-minute tab
# against a 20-minute video) is infeasible and librosa raises. Fall
# back to the unconstrained path rather than failing the whole sync.
_log.warning("gp_autosync: constrained DTW infeasible (%s) — "
"falling back to unconstrained steps", exc)
_D, wp = librosa.sequence.dtw(cs, ca, metric='cosine')
return wp[::-1] # reverse to forward order
# ── Sync point extraction from DTW path ──────────────────────────────────────