Eleven takes from the four 8.28 tapes, pulled apart twice over. First into six instrument
stems, keeping guitar and drums. Then the drum track again into six separate kit pieces.
11takes split
2stems kept each
6kit pieces each
24bit44.1 kHz wav
66kit lanes
What was asked
Separate the takes, keep one or two guitars and the drums, drop vocals, keys and bass.
Then run the drum track through a second separation so the kit arrives as individual pieces.
The measurements agree with the request. On every take
the bass stem came back near empty and the vocal and piano stems were effectively silent.
Nobody played bass or keys on any of these, so the lineup you asked for is the lineup
that was actually on the tape.
Why the mix was tilted first
The source files carry almost all of their energy in the low end. Song 13 measures
87.4% below 250 Hz and song 11 measures 92.6%, and every take on this page sits in the
same range, from 82.7% up to 96.7%. Each take's own figure is on its card below. That is
far outside the spectral balance Demucs was trained on, and the guitar head responds by
claiming low frequency content instead of guitar.
The tilt belongs ahead of separation, and the stems keep it. A 45 Hz high pass,
cuts of 9 dB at 90 Hz and 5 dB at 200 Hz, then lifts of 5 dB at 2.5 kHz and 7 dB at 6 kHz.
Separation then runs on the tilted audio.
Restoring the original tone afterwards undoes the whole thing, since the inverse boost
puts the low energy straight back. The delivered stems stay in the tilted domain and sound
brighter than the room. Add your own low shelf in the DAW if you want the tape sound back.
The check that proves it worked is the share of guitar energy above 500 Hz. Below about
15% means the pass failed. Song 13 came out at 44.1% and song 11 at 57.8%.
The tape is mono, so both guitars share one stem
Every tape records to a single mono channel. The usual
way to separate two guitarists takes Guitar 1 from the left channel and Guitar 2 from the
right, which needs a stereo field. On every file the left channel minus the right channel
measures exactly zero and the correlation reads 1.0. Both players sit inside one
guitar.wav. Splitting them by pitch range would be a guess dressed as a
measurement, so it was not done.
The eleven takes
10
10_tape1-59-59
2:03 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
82.7% below 250 Hz, 12.5% at 250-500,
4.5% at 500-2k, 0.3% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
48.85%
-7.1
99.6
0.3
0.1
0.0
snare
47.44%
-0.3
50.2
22.1
21.0
6.7
toms
0.26%
-12.8
77.7
13.8
8.0
0.5
hh
0.83%
-15.0
0.3
2.0
9.9
86.1
ride
0.19%
-17.2
0.1
1.9
6.0
86.0
crash
2.43%
-14.3
0.0
1.9
5.1
91.9
11
11_tape2-02-20
8:53 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
92.6% below 250 Hz, 4.7% at 250-500,
2.3% at 500-2k, 0.3% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
73.13%
-7.5
99.3
0.6
0.2
0.0
snare
23.78%
-0.3
47.7
21.1
23.5
7.5
toms
0.61%
-8.2
92.6
4.0
3.1
0.3
hh
0.48%
-12.2
0.3
2.5
15.6
80.0
ride
0.13%
-16.9
0.1
0.6
2.8
84.2
crash
1.87%
-12.8
0.0
1.0
8.0
90.1
13
13_tape2-21-16
10:34 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
87.4% below 250 Hz, 9.2% at 250-500,
3.0% at 500-2k, 0.5% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
86.59%
-7.7
99.6
0.3
0.1
0.0
snare
7.58%
-1.9
46.2
10.0
10.7
32.3
toms
0.42%
-16.4
96.6
1.6
1.4
0.3
hh
1.09%
-10.4
0.3
0.4
6.6
89.2
ride
0.25%
-15.5
0.1
0.2
1.8
86.6
crash
4.08%
-7.9
0.0
0.6
4.2
94.0
14
14_tape2-46-42
2:43 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
96.6% below 250 Hz, 2.6% at 250-500,
0.7% at 500-2k, 0.1% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
94.82%
-7.0
99.8
0.2
0.1
0.0
snare
2.32%
-15.8
38.4
13.4
20.8
26.5
toms
0.72%
-22.7
85.5
6.5
6.9
1.1
hh
1.23%
-24.7
0.2
1.4
13.6
80.5
ride
0.17%
-22.1
0.1
0.3
2.3
90.4
crash
0.74%
-24.9
0.0
0.9
12.9
84.8
15
15_tape2-52-01
8:02 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
96.7% below 250 Hz, 2.5% at 250-500,
0.6% at 500-2k, 0.1% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
92.09%
-7.0
99.3
0.5
0.2
0.0
snare
4.46%
-1.0
43.2
10.8
17.6
27.8
toms
1.57%
-15.9
98.0
0.8
1.0
0.2
hh
0.55%
-12.6
0.3
2.2
16.4
78.9
ride
0.17%
-13.9
0.1
0.4
3.5
90.9
crash
1.16%
-14.6
0.0
1.6
8.0
89.3
17
17_tape3-10-51
12:32 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
92.5% below 250 Hz, 4.6% at 250-500,
2.8% at 500-2k, 0.2% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
92.92%
-6.4
99.6
0.3
0.1
0.0
snare
4.12%
-3.0
42.6
12.0
18.5
26.2
toms
1.50%
-15.3
93.1
2.9
3.6
0.5
hh
0.53%
-17.5
0.4
3.0
20.5
74.6
ride
0.18%
-14.5
0.1
0.3
3.2
88.2
crash
0.75%
-13.4
0.0
1.0
11.5
86.5
18
18_tape3-23-56
6:10 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
89.1% below 250 Hz, 6.3% at 250-500,
3.3% at 500-2k, 1.2% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
42.44%
-7.4
99.8
0.1
0.0
0.0
snare
52.35%
-0.3
43.8
11.5
19.5
24.7
toms
0.12%
-19.9
86.8
6.0
4.7
2.5
hh
1.37%
-12.0
0.2
5.5
22.6
69.2
ride
0.05%
-28.7
0.3
1.6
2.5
75.6
crash
3.68%
-14.6
0.0
1.4
7.1
90.0
19
19_tape3-35-51
9:31 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
91.7% below 250 Hz, 7.4% at 250-500,
0.6% at 500-2k, 0.2% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
71.85%
-6.5
99.4
0.4
0.1
0.0
snare
19.29%
-7.5
40.6
9.3
23.4
26.1
toms
0.40%
-19.9
86.8
8.1
4.7
0.4
hh
1.87%
-16.2
0.5
1.3
19.5
74.2
ride
0.05%
-24.0
0.3
2.7
6.6
82.2
crash
6.55%
-12.9
0.0
1.6
9.0
88.4
20
20_tape3-47-38
3:15 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
98.9% below 250 Hz, 0.6% at 250-500,
0.3% at 500-2k, 0.2% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
92.84%
-7.8
99.7
0.2
0.1
0.0
snare
4.89%
-7.7
43.5
10.8
19.0
26.0
toms
0.85%
-19.7
90.4
4.9
4.5
0.3
hh
0.65%
-16.0
0.2
1.2
13.6
81.3
ride
0.12%
-21.4
0.1
0.6
4.1
87.6
crash
0.65%
-19.7
0.0
1.6
12.3
85.2
21
21_tape3-55-09
4:35 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
93.0% below 250 Hz, 5.6% at 250-500,
1.3% at 500-2k, 0.2% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
93.94%
-6.4
99.6
0.3
0.1
0.0
snare
3.58%
-7.6
41.0
10.3
18.4
29.8
toms
0.48%
-19.9
92.2
2.8
4.3
0.7
hh
0.87%
-14.3
0.2
2.1
19.4
74.7
ride
0.19%
-19.2
0.1
1.1
6.2
84.3
crash
0.93%
-19.4
0.0
0.9
9.6
88.0
22
22_tape4-00-08
4:34 · 44100 Hz · mono, max|L−R| = 0.0
Raw source spectrum, measured on the untouched file:
97.1% below 250 Hz, 2.2% at 250-500,
0.6% at 500-2k, 0.2% at 2k-8k.
Anything past 70% in that first band calls for the tilt.
Demucs, six stems
stem
energy
peak dB
<250
>500
Drum kit, six pieces
piece
energy
peak dB
<250
250-500
500-2k
2k-8k
kick
93.24%
-6.9
99.6
0.3
0.1
0.0
snare
3.70%
-11.4
50.3
10.0
15.8
23.3
toms
0.30%
-17.4
81.3
9.4
8.2
1.1
hh
0.45%
-20.5
0.3
2.2
19.0
77.2
ride
0.31%
-19.8
0.1
0.4
4.2
85.3
crash
2.00%
-14.8
0.0
1.2
12.0
85.6
The six-piece kit model
The drum split uses MDX23C DrumSep, the aufr33 and jarredou collaboration checkpoint.
The local copy of this tool was gone from the MacBook, so it was rebuilt from a
HuggingFace mirror and verified byte for byte at 437,652,699 bytes.
This mirror carries the six-stem model.
The earlier build merged ride and crash into a single cymbals lane. This one separates them
at the source, so ride.wav and crash.wav arrive already split and
no heuristic has to guess which cymbal was struck.
Where the files are
These are MacBook Pro paths. This work ran on the MacBook,
and the paths below resolve there. On Jacks-iMac they resolve to nothing, so a session
opening this page on the iMac will conclude the files are missing when they are not.
Checked from Jacks-iMac on 2026-09-06: the only complete six-lane kit reachable from that
machine is take 13, on the T7 Shield at
/Volumes/T7 Shield/New Folder With Items 3/New Folder With Items/song13/.
On the MacBook Pro, everything sits under
~/Projects/_outputs/band-practice-stems/8-28/, one folder per song.
band/ holds the two stems you asked to keep. drumkit/ holds the
six kit pieces. demucs/ keeps the full six-stem separation, including the
near-silent bass, vocal and piano stems, since that silence is the evidence for who was
playing.
Google Drive, read on 2026-09-06: gdrive:BandPractice/8-28-26-drumkit holds
take 11 only, 23 objects and 1.658 GiB, and roughly 635 MB of that is intermediate
scratch (_blocks/*.npy, .asd, .DS_Store) that should
not have shipped. The remaining ten takes have not been uploaded. A full audit of what is
verified where is at
drwu-htmls.vercel.app/band-practice-8-28-audit.