An earlier check said parts of a song's bass line might be off. This page compares the guessed bass notes from the music file with the actual bass sound from the recording, note by note. Green means they agree. Yellow means they mostly agree. Red means they disagree.
3 Gene · bass ear-check · 2026-07-07
Bass MIDI vs bass stem — flagged weak bars
Verifies the June 28 ear-check flags (bars 0, 37-39, 60-67, 84, 92-100) by comparing the FIXED MIDI's per-beat bass note against librosa.pyin pitch estimates from the isolated bass.mp3 stem.
Whole-song agreement
406
Total beats
373
w/ MIDI bass
373
w/ stem bass
373
Tested
248/373 = 0.665
Agreement
Flagged ranges
Range
Coverage
Agreement
bar 0
4 beats total, 0 tested
no tested beats
bars 37–39
12 beats total, 12 tested
9/12 = 0.75
bars 60–67
32 beats total, 32 tested
22/32 = 0.688
bar 84
4 beats total, 4 tested
3/4 = 0.75
bars 92–100
36 beats total, 36 tested
29/36 = 0.806
Top divergences per range
bar 0
no divergences (all tested beats agree, or nothing tested)
bars 37–39 · top divergences
Beat
Time
MIDI says
Stem says
Frames
37.4
100.525s
A
C#
28
38.4
103.184s
F#
D
28
39.4
105.819s
A
C#
28
bars 60–67 · top divergences
Beat
Time
MIDI says
Stem says
Frames
60.2
160.509s
C#
D
30
61.4
164.507s
A
C#
29
63.4
169.819s
A
C#
29
64.2
171.143s
D
D#
29
64.4
172.49s
F#
D
29
67.4
180.471s
A
C#
29
bar 84 · top divergences
Beat
Time
MIDI says
Stem says
Frames
84.2
224.526s
F#
E
18
bars 92–100 · top divergences
Beat
Time
MIDI says
Stem says
Frames
92.1
245.203s
E
F#
29
93.2
248.523s
D
F#
28
95.4
255.141s
A
B
28
96.2
256.465s
C#
D
28
95.2
253.817s
A
D
27
97.2
259.112s
A
D
27
Honest read
The bass FIXED MIDI is roughly 67% agreement with the isolated bass stem across the whole song, and the flagged ranges land at 69–81%. This is meaningfully better than random but not clean. The most common divergence pattern is the MIDI picks a chord-tone (root or 5th) while the stem is playing a passing tone (upper neighbor, chromatic approach) at the beat boundary. That looks like beat-window slop, not a chord misidentification. A re-transcription pass with a smaller window (half-beat) or a smarter voting rule would tighten agreement.
The bar 0 result (no tested beats) is expected: the MIDI intentionally has no bass in bar 0 per the fix report's count-in.
Not done this pass
basic-pitch rebuild of the bass MIDI from bass.mp3: not attempted. basic-pitch is not installed in this env. The librosa.pyin extract is a monophonic-only check, adequate for verifying the existing MIDI but not sufficient to fully rewrite it.
Re-patching FIXED.gp / FIXED.mid with corrections: deferred. This pass surfaces WHERE the disagreements are so a human can decide which to accept. Applying edits blindly at 66% agreement would risk introducing new errors.