Files 16 and 17, split
Two takes ran continuous with no stopping anywhere in them. The analysis found 13 places where the band changed songs, which turns 60 minutes into 15 tracks. Every cut carries 40 seconds of audio across it, so one listen settles it.
File 16, cut into 8
35:27 of continuous playing that starts at 2:30:17 on the tape. 7 places carried evidence of a song change.
Cut 1 of 7, thin evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -30 dB reads 144 into 144 bpm, which backs the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 2 of 7, strong evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -34 dB reads 152 into 136 bpm, which backs the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 3 of 7, strong evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -32 dB reads 144 into 136 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 4 of 7, thin evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -32 dB reads 136 into 123 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 5 of 7, strong evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -33 dB reads 136 into 118 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 6 of 7, likely evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -33 dB reads 129 into 129 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 7 of 7, thin evidence
20 seconds before the cut, 20 seconds after it.
Isolated bass at -19 dB reads 96 into 136 bpm, which backs the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
File 17, cut into 7
24:42 of continuous playing that starts at 3:06:47 on the tape. 6 places carried evidence of a song change.
Cut 1 of 6, likely evidence
20 seconds before the cut, 20 seconds after it.
Isolated bass at -16 dB reads 136 into 136 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 2 of 6, likely evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -24 dB reads 129 into 129 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 3 of 6, thin evidence
20 seconds before the cut, 20 seconds after it.
Isolated bass at -19 dB reads 112 into 118 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 4 of 6, strong evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -40 dB reads 144 into 152 bpm, which does not show the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 5 of 6, likely evidence
20 seconds before the cut, 20 seconds after it.
Isolated drums at -31 dB reads 136 into 96 bpm, which backs the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
Cut 6 of 6, likely evidence
20 seconds before the cut, 20 seconds after it.
Isolated bass at -15 dB reads 123 into 86 bpm, which backs the tempo step the mix suggested. Harmonic and timbral evidence is unaffected.
The 15 songs that came out
Sliced on the exact sample, so the pieces tile their parent with zero frames of
drift and the 24 bit values pass through untouched. The first pass used stream copy, which lands
on a packet edge and let neighbours overlap by up to 60 milliseconds. Those are parked in
songs-split/_packet_rounded_superseded/. The originals of 16 and 17 stay whole.
From file 16
2:55
30 seconds from the middle, enough to name it.
songs-split/16a_00-00_to_02-55.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 46.1% | 83% | -7.4 |
| bass | 31.6% | 93% | -13.0 |
| drums | 2.3% | 66% | -14.7 |
| vocals | 16.2% | 32% | -9.1 |
| other | 3.9% | 13% | -10.6 |
stems-split/16a_00-00_to_02-55/
5:55
30 seconds from the middle, enough to name it.
songs-split/16b_02-55_to_08-50.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 20.1% | 82% | -13.0 |
| bass | 40.9% | 96% | -13.9 |
| drums | 6.5% | 63% | -10.7 |
| vocals | 27.5% | 26% | -9.0 |
| other | 4.9% | 8% | -8.0 |
stems-split/16b_02-55_to_08-50/
5:04
30 seconds from the middle, enough to name it.
songs-split/16c_08-50_to_13-54.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 29.4% | 80% | -11.2 |
| bass | 58.8% | 82% | -11.3 |
| drums | 4.0% | 55% | -11.0 |
| vocals | 7.4% | 16% | -9.8 |
| other | 0.4% | 12% | -20.2 |
stems-split/16c_08-50_to_13-54/
4:41
30 seconds from the middle, enough to name it.
songs-split/16d_13-54_to_18-35.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 8.6% | 46% | -11.5 |
| bass | 66.0% | 90% | -12.7 |
| drums | 2.3% | 78% | -21.9 |
| vocals | 21.2% | 22% | -7.5 |
| other | 1.9% | 10% | -13.0 |
stems-split/16d_13-54_to_18-35/
6:46
30 seconds from the middle, enough to name it.
songs-split/16e_18-35_to_25-21.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 27.5% | 69% | -6.8 |
| bass | 13.8% | 62% | -12.2 |
| drums | 3.8% | 24% | -7.8 |
| vocals | 28.2% | 38% | -7.3 |
| other | 26.7% | 32% | -6.3 |
stems-split/16e_18-35_to_25-21/
3:20
30 seconds from the middle, enough to name it.
songs-split/16f_25-21_to_28-41.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 26.0% | 72% | -15.2 |
| bass | 68.5% | 46% | -13.3 |
| drums | 4.6% | 87% | -23.3 |
| vocals | 0.0% | 6% | -35.2 |
| other | 0.8% | 14% | -18.9 |
stems-split/16f_25-21_to_28-41/
2:59
30 seconds from the middle, enough to name it.
songs-split/16g_28-41_to_31-40.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 42.5% | 92% | -15.8 |
| bass | 57.1% | 61% | -13.5 |
| drums | 0.4% | 15% | -23.7 |
| vocals | 0.0% | 3% | -35.7 |
| other | 0.0% | 100% | -43.9 |
stems-split/16g_28-41_to_31-40/
3:47
30 seconds from the middle, enough to name it.
songs-split/16h_31-40_to_35-27.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 27.9% | 90% | -15.5 |
| bass | 72.1% | 94% | -13.1 |
| drums | 0.0% | 100% | -37.1 |
| vocals | 0.0% | 100% | -58.6 |
| other | 0.0% | 100% | -37.4 |
stems-split/16h_31-40_to_35-27/
From file 17
4:17
30 seconds from the middle, enough to name it.
songs-split/17a_00-00_to_04-17.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 2.8% | 22% | -15.1 |
| bass | 77.2% | 97% | -12.6 |
| drums | 0.0% | 5% | -35.4 |
| vocals | 19.9% | 28% | -8.9 |
| other | 0.1% | 9% | -26.5 |
stems-split/17a_00-00_to_04-17/
4:54
30 seconds from the middle, enough to name it.
songs-split/17b_04-17_to_09-11.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 2.1% | 32% | -16.9 |
| bass | 83.2% | 98% | -11.9 |
| drums | 0.2% | 14% | -25.1 |
| vocals | 14.1% | 12% | -8.7 |
| other | 0.4% | 9% | -18.2 |
stems-split/17b_04-17_to_09-11/
3:17
30 seconds from the middle, enough to name it.
songs-split/17c_09-11_to_12-28.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 33.0% | 39% | -8.3 |
| bass | 18.6% | 98% | -14.1 |
| drums | 29.9% | 62% | -6.6 |
| vocals | 0.2% | 4% | -17.7 |
| other | 18.3% | 63% | -9.2 |
stems-split/17c_09-11_to_12-28/
2:36
30 seconds from the middle, enough to name it.
songs-split/17d_12-28_to_15-04.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 60.6% | 91% | -7.1 |
| bass | 12.5% | 87% | -13.5 |
| drums | 0.0% | 16% | -33.0 |
| vocals | 9.2% | 29% | -9.3 |
| other | 17.7% | 32% | -6.9 |
stems-split/17d_12-28_to_15-04/
3:10
30 seconds from the middle, enough to name it.
songs-split/17e_15-04_to_18-14.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 2.6% | 63% | -13.3 |
| bass | 30.1% | 90% | -13.5 |
| drums | 0.4% | 71% | -25.1 |
| vocals | 53.0% | 29% | -6.1 |
| other | 13.9% | 21% | -8.9 |
stems-split/17e_15-04_to_18-14/
3:47
30 seconds from the middle, enough to name it.
songs-split/17f_18-14_to_22-01.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 10.2% | 81% | -14.2 |
| bass | 43.9% | 98% | -12.9 |
| drums | 3.0% | 44% | -12.5 |
| vocals | 41.2% | 28% | -7.0 |
| other | 1.6% | 7% | -10.3 |
stems-split/17f_18-14_to_22-01/
2:41
30 seconds from the middle, enough to name it.
songs-split/17g_22-01_to_24-42.wav
five stems, measured
| stem | energy share | playing | peak dB |
|---|---|---|---|
| guitar | 7.3% | 57% | -14.7 |
| bass | 51.2% | 98% | -13.1 |
| drums | 2.0% | 7% | -13.7 |
| vocals | 36.4% | 24% | -7.7 |
| other | 3.2% | 7% | -11.2 |
stems-split/17g_22-01_to_24-42/
Stems for the whole set
The tape is mono and it carries most of its energy in the low end, which sits outside what Demucs was trained on. Separating it raw puts the guitar stem down in bass register, measured at 93.9 percent below 250 Hz on song 04 in the earlier pass. So the mix gets tilted before separation, and the inverse tilt goes back on every delivered stem so the tone matches the source.
Two models run on each song. htdemucs_6s carries the only guitar head, so guitar and other come from there. htdemucs holds more capacity per source across four heads, so bass, drums and vocals come from that one. Every stem sums to one channel, since the source channels are bit identical. Five stems per song at 44.1 kHz 24 bit.
The drum chair rotates through this band. A silent drums stem on a take means an empty chair, and it is not a separation failure. Songs 04, 05 and 06 already came through this pipeline last night and their numbers are on the stems page.
The set is 32 songs and 148 minutes, running as two parallel chains.
Output lands in ~/Projects/_outputs/band-practice-stems/STEMS/, one folder per song.
How a cut gets found
Four signals vote on every second of a take. Chroma self similarity catches a new tonal center. MFCC self similarity catches a new arrangement. Local tempo catches a step in the pulse that holds. Short level dips catch the breath between pieces. A checkerboard kernel sized in minutes turns the four into one novelty curve, since a song boundary is a block change and a chorus is not. Peaks land at least 150 seconds apart, then snap to the quietest instant within ten seconds so the cut falls in a breath.
The number on each card is that combined score. Above 0.55 reads strong, above 0.40 reads likely, below that reads thin and wants your ear.