Compress to Create - Jean-Pierre Briot - MIDI Files

Note that unfortunately some of the original midi files were lost (overwritten). We then substitute them with some recent generation, when indicated - other generation.

Figure 4: "The Green Mountain" (first 8 measures)

Figure 5: "Willa Fjord" (first 8 measures)

Figure 8: Example of melody generated from a random latent vector by the decoder component of the autoencoder (h = 1500) trained on the Celtic melodies corpus - other generation

Figure 11: (from top to bottom) Reconstruction by the autoencoder of "The Green Mountain" for h = 1000, 750, 500, 250, 200, 150, 100, 2

• h = 1000
• h = 750
• h = 500
• h = 250
• h = 200
• h = 150
• h = 100
• h = 2

Figure 13: (from top to bottom) Melodies resulting from the interpolation (5 steps) by the autoencoder (h = 1500), from "The Green Mountain" to "Willa Fjord"

• step 0: Start: The Green Mountain
• step 1
• step 2
• step 3
• step 4
• step 5: Target: Willa Fjord

Figure 15: Melody resulting from the interpolation (5 steps) by the autoencoder (h = 2) of the value of z1 (from its min value to its max value), while z2 is constantly equal to its mean value - other generation

Figure 16: Melody resulting from the interpolation (5 steps) by the autoencoder (h = 2) of the value of z2 (from its min value to its max value), while z1 is constantly equal to its mean value - other generation

Figure 17: Bach Chorale BWV 347 soprano voice (transposed into D Major key) first 8 measures

Figure 19: "The Green Mountain" transformed into a Bach chorales-like melody by the autoencoder (h = 1500)

Figure 21: Bach Chorale BWV 347 soprano voice transformed into a Celtic-like melody by the autoencoder (h = 1500) - other generation

Figure 24: (from top to bottom) Example of successive melodies generated by the autoencoder (h = 1500) for each step of the recursion - other generation

• step 0: Start: Random Seed
• step 1
• step 2
• step 3
• step 4: Fixed Point

Figure 26: Example of melody generated by the autoencoder (h = 1500) with the objective of its first note being a C4 - other generation

Figure 27: Example of melody generated by the autoencoder (h = 1500) with the objective of maximizing the number of hold - other generation

Figure 28: Example of melody generated by the autoencoder (h = 1500) with the objective of minimizing the number of hold - other generation

Figure 29: (from top to bottom) Melodies resulting from the interpolation (5 steps) by the variational autoencoder (h = 1500), from "The Green Mountain" to "Willa Fjord" - other generation

• step 0: Start: The Green Mountain
• step 1
• step 2
• step 3
• step 4
• step 5: Target: Willa Fjord

Figure 30: (from top to bottom) Melodies resulting from the interpolation (5 steps) by the variational autoencoder (h = 2) of the value of z1 (from its min value to its max value), while z2 is constantly equal to its mean value

• step 0: Start: z1 = min(z1)
• step 1
• step 2
• step 3
• step 4
• step 5: Target: z1 = max(z1)

Figure 31: (from top to bottom) Melodies resulting from the interpolation (5 steps) by the variational autoencoder (h = 2) of the value of z2 (from its min value to its max value), while z1 is constantly equal to its mean value

• step 0: Start: z2 = min(z2)
• step 1
• step 2
• step 3
• step 4
• step 5: Target: z2 = max(z2)

Figure 32: "The Green Mountain" transformed into a Bach chorales-like melody by the variational autoencoder (h = 1500)

Figure 33: Example of melody generated by the variational autoencoder (h = 1500) by recursion

Figure 34: Example of melody generated by the variational autoencoder (h = 1500) with the objective of its first note being a C4

Figure 35: Example of melody generated by the variational autoencoder (h = 1500) with the objective of maximizing the number of hold

Figure 36: Example of melody generated by the variational autoencoder (h = 1500) with the objective of minimizing the number of hold