Skip to main content

Table 7 Correlation coefficients between the MUSHRA scores and the L1 and MR-STFT errors for the conditions of the MUSHRA listening test. In the last row (“Total”), the samples of the different conditions are considered together. \(\mathcal {L}\) is the sum of the L1 and MR-STFT errors and is the loss function used for training HT Demucs and the GCNTF networks

From: Blind extraction of guitar effects through blind system inversion and neural guitar effect modeling

 

Softclipping

Phaser

Slapback delay

 

L1

MR-STFT

\(\mathcal {L}\)

L1

MR-STFT

\(\mathcal {L}\)

L1

MR-STFT

\(\mathcal {L}\)

\(NN_{200}\)

− 0.308

− 0.508

− 0.49

− 0.908

− 0.551

− 0.834

− 0.327

− 0.453

− 0.435

\(NN_{500}\)

− 0.668

− 0.804

− 0.794

− 0.833

− 0.813

− 0.825

0.312

− 0.51

− 0.165

\(NN_{1000}\)

− 0.8

− 0.788

− 0.797

− 0.155

0.379

0.25

− 0.462

− 0.583

− 0.579

Total

− 0.693

− 0.789

− 0.784

0.035

− 0.738

− 0.658

− 0.021

− 0.518

− 0.38