SPTVMod - Fast-LFO Dataset Spectrograms

This page presents the output spectrograms of the models trained in Experiment 2 using the 4 s long chirp-train input signal. Multiple models were trained for each combination of mode seeking hyperparameters (λ, ε) and each dataset. We inspect the results obtained after the adversarial phase, then after the fine-tuning phase.

λ = 0 means that mode seeking is disabled.

The spectrograms use boxcar chirp-aligned windows, i.e. the window and hop sizes are equal to the chirp length (1336 samples); the frequency axis ranges from 0.04 to 17 kHz in log-scale.

Model	Adversarial Phase Checkpoint: best modulation loss	Spectral Fine-tuning (no SPN pre-training) Checkpoint: best modulation loss	Spectral Fine-tuning (no SPN pre-training) Checkpoint: best spectral loss	Spectral Fine-tuning (with SPN pre-training) Checkpoint: best modulation loss	Spectral Fine-tuning (with SPN pre-training) Checkpoint: best spectral loss
1 λ=0.01 ε=0.001
2 λ=0.01 ε=0.01
3 λ=0.01 ε=0.1
4 λ=0.1 ε=0.001
5 λ=0.1 ε=0.01
6 λ=0.1 ε=0.1
7 λ=1.0 ε=0.001
8 λ=1.0 ε=0.01
9 λ=1.0 ε=0.1
10 λ=0
11 λ=0.01 ε=0.01
12 λ=0
13 λ=0.01 ε=0.001
14 λ=0.01 ε=0.01
15 λ=0.01 ε=0.1
16 λ=0.1 ε=0.001
17 λ=0.1 ε=0.01
18 λ=0.1 ε=0.1
19 λ=1.0 ε=0.001
20 λ=1.0 ε=0.01
21 λ=1.0 ε=0.1
22 λ=0
23 λ=0.01 ε=0.01
24 λ=0
25 λ=0.01 ε=0.001
26 λ=0.01 ε=0.01
27 λ=0.01 ε=0.1
28 λ=0.1 ε=0.001
29 λ=0.1 ε=0.01
30 λ=0.1 ε=0.1
31 λ=1.0 ε=0.001
32 λ=1.0 ε=0.01
33 λ=1.0 ε=0.1
34 λ=0
35 λ=0.01 ε=0.01
36 λ=0
37 λ=0
38 λ=0
39 λ=0
40 λ=0
41 λ=0
42 λ=0

Reference (Target Effect's Output)