Every keeper take from the session, pulled apart into separate instrument tracks. The tape is mono and extremely bass heavy, and both of those facts decide what can be recovered here.
19of 19 songs
95stem files
5stems per song
1guitar stem, not 2
24bit44.1 kHz wav
What mono costs
There is one guitar stem, not two. The usual way to separate two guitarists is to take Guitar 1 from the left channel and Guitar 2 from the right. That needs a stereo field. On all three of these files the left channel minus the right channel is exactly zero and the correlation reads 1.0, so the two channels are bit-identical. Both players sit inside one guitar.wav. Splitting them by pitch range would be a guess dressed up as a measurement, so it was not done.
Two questions a mono practice recording cannot settle, and both are marked untested rather than upgraded: which guitarist played a given line, and any claim about stereo placement. Everything else on this page is measured.
The guitar stem fix
A first pass produced a guitar stem that sat in bass register. Measured on song 04, the guitar track held 88.7% of its energy below 250 Hz and 3.1% above 500 Hz. That is bass register, not guitar register.
The cause was the tape, not the model. These recordings carry roughly 88% of their energy below 200 Hz, far outside the spectral balance Demucs was trained on. The guitar head was handed almost nothing above 250 Hz to work with, so it took bass register instead.
The fix is to tilt the mix toward a normal balance ahead of separation: a 45 Hz high pass, then 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 runs on the tilted audio.
song 04 guitar stem
below 250 Hz
above 500 Hz
first pass, no tilt
88.7%
3.1%
tilt, then inverse tilt on output
90.9%
3.0%
tilt, no inverse
53.8%
33.7%
Restoring the original tone afterwards defeats the whole thing, since the inverse boost puts the low energy straight back. The delivered stems therefore stay in the tilted domain and sound brighter than the room. Add your own low shelf downstream if you want the tape sound back.
The songs
Who was playing decides the drums row. The drum chair rotates, and on some takes it goes empty because that player moved to second guitar or to bass instead. The measurements match that: take 32 carries a kit at 19.8% of energy peaking at -13.8 dB, and takes 36 and 41 carry nothing above -40 dB. Both models agree on all three songs, so the silence is an empty drum chair rather than a separation failure.
both guitarists, summed. Mono gives no pan cue to split them.
bass
17.7%
74%
-18.3
from the 4-stem model, which separates low end better.
drums
0.0%
6%
-54.6
from the 4-stem model.
vocals
0.0%
9%
-38.9
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.2%
25%
-30.8
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 4:07, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
11.3%
27%
-16.5
from the 4-stem model, which separates low end better.
drums
11.0%
54%
-22.0
from the 4-stem model.
vocals
0.0%
8%
-53.0
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.1%
23%
-40.3
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 2:27, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
1.8%
29%
-21.7
from the 4-stem model, which separates low end better.
drums
8.0%
89%
-20.9
from the 4-stem model.
vocals
0.0%
2%
-43.9
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.0%
8%
-51.4
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 3:42, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
9.3%
79%
-20.8
from the 4-stem model, which separates low end better.
drums
15.7%
92%
-20.2
from the 4-stem model.
vocals
0.0%
12%
-49.2
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.0%
25%
-56.1
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 3:39, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
10.2%
40%
-18.8
from the 4-stem model, which separates low end better.
drums
0.0%
38%
-59.9
from the 4-stem model.
vocals
0.0%
6%
-48.2
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.2%
5%
-29.1
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 5:03, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
8.9%
35%
-18.4
from the 4-stem model, which separates low end better.
drums
0.0%
6%
-50.5
from the 4-stem model.
vocals
0.9%
3%
-20.9
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.1%
4%
-29.2
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 10:54, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
8.7%
58%
-22.5
from the 4-stem model, which separates low end better.
drums
0.0%
100%
-70.8
from the 4-stem model.
vocals
0.0%
1%
-41.4
from the 4-stem model. Silent stretches are the instrumental passes.
other
1.6%
11%
-26.3
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 4:20, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
29.3%
70%
-15.7
from the 4-stem model, which separates low end better.
drums
0.0%
2%
-38.8
from the 4-stem model.
vocals
0.1%
5%
-27.1
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.0%
22%
-54.8
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 4:35, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
23.9%
77%
-16.6
from the 4-stem model, which separates low end better.
drums
0.0%
3%
-49.7
from the 4-stem model.
vocals
0.4%
4%
-23.1
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.0%
4%
-34.4
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 5:01, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
2.3%
21%
-21.4
from the 4-stem model, which separates low end better.
drums
16.9%
77%
-19.2
from the 4-stem model.
vocals
4.1%
4%
-12.2
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.8%
2%
-20.6
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 4:58, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
7.2%
34%
-19.5
from the 4-stem model, which separates low end better.
drums
7.7%
53%
-20.8
from the 4-stem model.
vocals
37.1%
13%
-6.6
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.3%
5%
-25.8
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 9:38, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
17.9%
46%
-12.6
from the 4-stem model, which separates low end better.
drums
4.5%
21%
-14.5
from the 4-stem model.
vocals
27.0%
9%
-7.1
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.0%
15%
-41.2
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 9:16, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
65.9%
52%
-14.4
from the 4-stem model, which separates low end better.
drums
0.2%
0%
-21.9
from the 4-stem model.
vocals
13.8%
3%
-7.5
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.0%
100%
-44.8
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 2:27, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
80.7%
88%
-13.8
from the 4-stem model, which separates low end better.
drums
0.0%
100%
-57.6
from the 4-stem model.
vocals
0.0%
100%
-55.8
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.6%
8%
-22.2
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 3:12, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
58.6%
44%
-13.9
from the 4-stem model, which separates low end better.
drums
0.0%
0%
-26.9
from the 4-stem model.
vocals
3.1%
6%
-16.1
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.1%
1%
-21.6
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 8:19, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
39.5%
78%
-11.3
from the 4-stem model, which separates low end better.
drums
3.6%
33%
-7.8
from the 4-stem model.
vocals
19.6%
19%
-7.3
from the 4-stem model. Silent stretches are the instrumental passes.
other
10.5%
10%
-6.3
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 35:27, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
41.1%
96%
-11.9
from the 4-stem model, which separates low end better.
drums
6.2%
14%
-6.6
from the 4-stem model.
vocals
23.0%
20%
-6.1
from the 4-stem model. Silent stretches are the instrumental passes.
other
9.1%
16%
-6.9
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 24:42, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
10.6%
46%
-13.7
from the 4-stem model, which separates low end better.
drums
5.2%
6%
-7.9
from the 4-stem model.
vocals
56.8%
20%
-6.1
from the 4-stem model. Silent stretches are the instrumental passes.
other
13.3%
17%
-6.9
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 2:12, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
both guitarists, summed. Mono gives no pan cue to split them.
bass
8.7%
38%
-14.5
from the 4-stem model, which separates low end better.
drums
1.3%
10%
-16.3
from the 4-stem model.
vocals
0.0%
47%
-56.2
from the 4-stem model. Silent stretches are the instrumental passes.
other
0.5%
4%
-19.5
everything the six heads could not claim.
guitar
bass
drums
vocals
other
Loudness envelope across the full 3:50, half second windows. The faint line is that stem's own activity gate, 25 dB below its peak. Flat stretches below the line are where the instrument is not playing.
Pulling one guitar out, and tuning it
Your band records to one mono channel, so there is no pan cue and no ground-truth way to split two guitarists into two audio files. What runs instead is a Viterbi melody tracker over the mixed guitar stem: per frame it takes the top autocorrelation peaks as candidates, then finds the smoothest path that rewards salience and pitch height, since the lead sits on top, and penalises large jumps. That follows one continuous voice through the mix.
Why the naive approach fails. A per-frame loudest-pitch tracker bounces between the two guitarists. Measured on song 01 over the same 40 seconds: the naive tracker moved 11.94 semitones at the 90th percentile with 14.3% of jumps over 7 semitones, spanning MIDI 40 to 85. The Viterbi path moved 0.12 semitones at the 90th percentile with 0.9% large jumps, spanning MIDI 52 to 77.
A harmonic comb mask around the tracked pitch then splits the audio into isolated.wav, the tracked lead, and residual.wav, everything else including the other guitarist. The two sum back to the original to machine precision, worst case 1.3e-05 RMS error across all 19 songs, so the mask is an exact partition rather than a lossy guess. Pitch correction is TD-PSOLA, snapping each voiced frame to the nearest equal-tempered semitone.
song
cents error, raw
cents error, corrected
within 10 cents, raw
within 10 cents, corrected
01
8.3
2.1
54.6%
86.8%
02
13.4
5.5
39.3%
63.8%
03
12.1
5.8
42.2%
63.6%
04
10.5
6.4
48.2%
64.2%
05
8.6
3.8
55.6%
76.6%
06
9.1
4.0
54.1%
75.5%
07
8.2
3.4
57.2%
81.3%
08
6.9
3.7
62.4%
79.2%
09
7.1
3.0
61.4%
83.5%
10
8.6
2.7
57.0%
84.5%
11
11.2
4.8
46.4%
71.5%
12
13.0
5.0
41.0%
69.4%
13
7.1
3.1
64.1%
81.9%
14
7.0
4.0
61.3%
76.4%
15
7.0
3.0
62.5%
81.6%
16
13.9
6.1
39.5%
64.1%
17
13.8
6.3
38.5%
62.5%
18
13.0
5.0
40.9%
67.0%
19
15.4
6.5
36.1%
64.1%
Improved on 19 of 19 songs, worsened on 0.
Two limits worth stating. This is not a true two-guitar audio split. Where both players share a harmonic at the same instant, a mono recording cannot resolve them and some bleed stays in the isolated track. And note transitions are the weak spot: individual frames at a change of note can overshoot the target by 40 cents or more, where the pitch tracker is least stable, even while the whole-song numbers move strongly in the right direction.
How it was separated
Every song went through two Demucs v4 passes, and the final stem set takes the best source from each.
model
gives
used for
htdemucs_6s
drums, bass, other, vocals, guitar, piano
the guitar stem, since this is the only Demucs model carrying a guitar head.
htdemucs
drums, bass, other, vocals
bass, drums and vocals. Four heads means more capacity per source, so the low end and the vocal come out cleaner.
Both ran on CPU at 24 bit to match the source depth. The sources are bit-identical across channels, and Demucs still writes two channels back out, because the model runs a stereo path internally. Those two output channels differ slightly: the guitar stem on take 41 measured a maximum channel difference of 0.0136, about 37 dB below full scale. Each delivered stem is the average of the two, which reconstructs the mono source and averages down the uncorrelated part of the model artifacts. Every stem records its own measured difference in manifest.json.
Reading the piano row. This band has no piano. Whatever energy lands in that stem is material the six-head model could not file correctly, which on a room recording is usually guitar upper harmonics and cymbal wash. A high piano share is a warning about that song's separation quality, not a discovery.
Sources are the keeper exports from the 8.14.26 practice tape, at ~/Projects/_outputs/real-song-finder/band-practice-8-14-26/songs/. The take selection that produced them is documented on the keepers page.