Interactive Figures for "Leveraging Neural Representations for Audio Manipulation"

Interactive versions of figures appearing in the paper by Hawley & Steinmetz, AES Europe, 2023

Created on April 9|Last edited on April 18

Comment

Preprint arXiv.2304.04394: Abstract, PDF﻿
Note that the 3D figures are made with Plotly which only supports ~10  figures at a time. 
Figures and Data
Sample Audio Effects (e.g., Figure 1 in the paper)These are sampled from GuitarSet, effects performed via Pedalboard.
﻿
runs.summary["audio_specs/table"]
 - 10 of 10
audio
melspec
embspec_stacked
1
2
3
4
5
6
7
8
9
10
Run set1
﻿
﻿
3D Plots of Audio Effects Classes (Figures 3 and 4 in the paper)"_dvae_" refers to the DiffAE model, and "_stacked_" refers to the Stacked DiffAE model.
Note: Varying the UMAP parameters (n_neighbors, min_dist) does not significantly alter the character of the plots. 
﻿
Run set1
﻿
﻿
3D Plots of Effect Parameter Paths (Figures 6 through 9 in the paper)For time-averaged representations. "_dvae_" refers to the DiffAE model, and "_stacked_" refers to the Stacked DiffAE model. PCA preserves straight lines (unlike UMAP), so these trajectories are truly curved.
﻿
Run set1
﻿
﻿

Add a comment