Comparison of neural network architectures for guitar amplifier emulation

Architecture	RTF	MSE	ESR	MAE	STFT	Target	Target + Cabinet Simulation	Prediction	Prediction + Cabinet Simulation	Reference
RNN (LSTM-32)	0.51	0.0040	0.0244	0.0378	0.5952					Wright, A.; Damskagg, E.P.; Valimaki, V. Real-Time Black-Box Modelling With Recurrent Neural Networks. Proceedings of the 22nd International Conference on Digital Audio Effects (DAFx-19) 2019, pp. 1-8. http://dafx.de/paper-archive/2019/DAFx2019_paper_43.pdf.
CNN (WaveNet)	0.35	0.0703	0.4337	0.1359	0.6542					Damskagg, E.p.; Juvela, L.; Thuillier, E.; Valimaki, V. Deep Learning for Tube Amplifier Emulation 2019. pp. 471-475. https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8682805.
Hybrid (conv-LSTM)	3.25	0.0069	0.0423	0.0530	0.6937					Schmitz, T. Nonlinear Modeling of the Guitar Signal Chain Enabling its Real-time Emulation 2019. p. 258. https://pdfs.semanticscholar.org/18e8/0acdd9d704a61a1f174a2a4a1a9411801785.pdf, doi:10.3115/v1/w14-4012.
CNN (Shallow TCN)	0.14	0.3190	2.1371	0.4510	1.2348					Steinmetz, C.J.; Reiss, J.D. Efficient Neural Networks for Real-time Analog Audio Effect Modeling 2021. [2102.06200]. http://arxiv.org/abs/2102.06200.

• Cabinet Simulation:
  DynIR Capture made by Two notes from a Two notes® Custom CSG Guitar cabinet.
  The capture was made on a 1x12 Open-back cabinet loaded with Electro-voice® 12" EVM12L speakers.
  Source: https://www.two-notes.com/en/product/2notes-csg/
• All targets and predictions (both with and without the cabinet simulation) were rendered using Reaper with volumes: Master at -0.85 dB and track at -5.23 dB.