roygbiv33 commited on
Commit
cf9e3e5
Β·
verified Β·
1 Parent(s): 935c1c9

Examples: elegy only (before/after pair)

Browse files
.gitattributes CHANGED
@@ -4,3 +4,5 @@
4
  output_examples/elegy_in_c_sharp_minor_performed.mp3 filter=lfs diff=lfs merge=lfs -text
5
  output_examples/fantaisie_01_score_as_composed.mp3 filter=lfs diff=lfs merge=lfs -text
6
  output_examples/fantaisie_02_performed.mp3 filter=lfs diff=lfs merge=lfs -text
 
 
 
4
  output_examples/elegy_in_c_sharp_minor_performed.mp3 filter=lfs diff=lfs merge=lfs -text
5
  output_examples/fantaisie_01_score_as_composed.mp3 filter=lfs diff=lfs merge=lfs -text
6
  output_examples/fantaisie_02_performed.mp3 filter=lfs diff=lfs merge=lfs -text
7
+ output_examples/elegy_01_score_as_composed.mp3 filter=lfs diff=lfs merge=lfs -text
8
+ output_examples/elegy_02_performed.mp3 filter=lfs diff=lfs merge=lfs -text
output_examples/README.md CHANGED
@@ -1,37 +1,27 @@
1
  # Output examples
2
 
3
- Every note in these files was composed by the model. The pairs below contain **identical pitches** β€”
4
- nothing added, removed or re-pitched β€” so the only thing that changes between them is *how they are
5
- played*.
6
 
7
- ## Fantaisie (F# major)
 
8
 
9
  | file | |
10
  |---|---|
11
- | `fantaisie_01_score_as_composed.mid` / `.mp3` | raw model output: no pedal, near-flat velocities, onsets square on the beat |
12
- | `fantaisie_02_performed.mid` / `.mp3` | the same 829 notes performed β€” melody voiced above the accompaniment, chords rolled by choice, rubato around a felt pulse, 174 pedal events caught at harmony changes |
13
 
14
- ## Elegy in C# minor
15
 
16
- | file | |
17
- |---|---|
18
- | `elegy_score_as_composed.mid` | raw model output |
19
- | `elegy_performed.mid` / `elegy_in_c_sharp_minor_performed.mp3` | the same 608 notes performed (196 pedal events) |
20
-
21
- Both were composed plan-driven: the model states a theme, develops it (fragmenting, inverting,
22
- transposing, augmenting) and brings it back transformed.
23
-
24
- ## Why the pairs matter
25
-
26
- Play the fantaisie's `01` and then `02`. The gap between them is the most useful thing we learned
27
- building this model: **a dry-sounding symbolic generation is usually not a composition failure but an
28
- unperformed one.** Pitch content and performance are separable layers, and the performance layer
29
- carries far more of the perceived quality than we expected.
30
 
31
  There's an acoustic reason as well as a musical one. Staggering a chord's attacks by roughly 20–70 ms
32
  spreads its beating partials out in time, which measurably lowers sensory roughness, so rolled chords
33
  aren't merely more expressive than struck ones, they're *cleaner*. Quantising a dense passage onto a
34
  grid makes it rougher, not just stiffer.
35
 
36
- The practical implication: if you're evaluating this model (or any symbolic music model) on raw
37
- output, you're hearing about half of what's there.
 
1
  # Output examples
2
 
3
+ **Elegy in C-sharp minor** β€” 608 notes, every one composed by the model (plan-driven: it states a
4
+ theme, develops it by fragmenting, inverting and transposing, then brings it back transformed).
 
5
 
6
+ Both files below contain **identical pitches**. Nothing was added, removed or re-pitched. The only
7
+ difference is *how they are played*.
8
 
9
  | file | |
10
  |---|---|
11
+ | `elegy_01_score_as_composed.mid` / `.mp3` | raw model output: no pedal, near-flat velocities, onsets square on the beat |
12
+ | `elegy_02_performed.mid` / `.mp3` | the same 608 notes performed β€” melody voiced above the accompaniment, chords rolled by choice, rubato around a felt pulse, 196 pedal events caught at harmony changes |
13
 
14
+ ## Why the pair matters
15
 
16
+ Play `01` and then `02`. The gap between them is the most useful thing we learned building this
17
+ model: **a dry-sounding symbolic generation is usually not a composition failure but an unperformed
18
+ one.** Pitch content and performance are separable layers, and performance carries far more of the
19
+ perceived quality than we expected.
 
 
 
 
 
 
 
 
 
 
20
 
21
  There's an acoustic reason as well as a musical one. Staggering a chord's attacks by roughly 20–70 ms
22
  spreads its beating partials out in time, which measurably lowers sensory roughness, so rolled chords
23
  aren't merely more expressive than struck ones, they're *cleaner*. Quantising a dense passage onto a
24
  grid makes it rougher, not just stiffer.
25
 
26
+ If you're evaluating this model (or any symbolic music model) on raw output, you're hearing about
27
+ half of what's there.
output_examples/{fantaisie_01_score_as_composed.mp3 β†’ elegy_01_score_as_composed.mp3} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a92af0abcb02dd595bf625270003ca9e15c4ad10e027fc33ef534ef67d11c2ab
3
- size 2367364
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:217188f3a7b79ebec5ff10e2cd5f292cbadb539881422826cb00348ee5b92b55
3
+ size 2846973
output_examples/{fantaisie_02_performed.mp3 β†’ elegy_02_performed.mp3} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4a1afe12b45813ceb8607b2176f34c8397c9acc024ddb6f714b95808a320cf26
3
- size 2954806
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2625d4faef31573c9289c8739a5f331f87c5b73d52daf6fdc6d8dd1062b00f9
3
+ size 3145395