NAM LSTM Experimentations

For configuration please go to NAM LSTM Configuration

Full Rig - Micro Cube GX (SM57)

Original

NAM (WaveNet, Nano) NAM file 01.78% CPU, ESR: 0.0110, 580 epochs Training stagnated past ~350 epochs
Only minimal difference in frequency response/curve to original. Far superior in this full_rig comparison.

NAM (LSTM) NAM file num_layers=2 hidden_size=21 01.5% CPU, ESR: 0.01067, 787 epochs
Highs are lost and it is not able to transparently recreate the frequency response. E.g. the narrow cut around 5000 Hz is lost. I need to further tweak configuration to hopefully remedy this.

CNLA 3972 Leveler (~20 dB Attenuation at -0.5 dBFS)

Original

NAM (WaveNet, Standard) NAM file 03.54% CPU, ESR: 0.001119, 1303 epochs

NAM (WaveNet, Nano) NAM file 01.78% CPU, ESR: 0.002329, 1153 epochs

NAM (LSTM) NAM file num_layers=2 hidden_size=14 00.99% CPU, ESR: 0.000919, 928 epochs

High Gain Pedal v2

Original

NAM (WaveNet, Standard) NAM file 03.54% CPU, ESR: 0.004417, 1050 epochs

NAM (WaveNet, Nano) NAM file 01.78% CPU, ESR: 0.02961, 670 epochs Training stagnated past 300 epochs

NAM (LSTM, NEW configuration) NAM file num_layers=2 hidden_size=14 00.99% CPU, ESR: 0.00439, 692 epochs

NAM (LSTM, OLD configuration) NAM file num_layers=2 hidden_size=14 00.99% CPU, ESR: 0.00656, 625 epochs

NAM (LSTM) NAM file num_layers=2 hidden_size=8 00.52% CPU, ESR: 0.01072, 729 epochs

NAM (LSTM) NAM file num_layers=1 hidden_size=8 00.32% CPU, ESR: 0.06119, 750 epochs

High Gain Pedal

Original

NAM (WaveNet, Nano) NAM file 01.78% CPU, ESR: 0.0799, 482 epochs

NAM (LSTM) NAM file num_layers=2 hidden_size=14 00.99% CPU, ESR: 0.0080, 483 epochs

NAM (LSTM) NAM file num_layers=2 hidden_size=8 00.52% CPU, ESR: 0.0097, 482 epochs

NAM (LSTM) NAM file num_layers=1 hidden_size=8 00.32% CPU, ESR: 0.0614, 320 epochs

Tube Mic Preamp

Original

NAM (WaveNet, Nano) NAM file 01.78% CPU, ESR: 0.0075, 420 epochs

NAM (LSTM) NAM file num_layers=2 hidden_size=14 00.99% CPU, ESR: 0.0020, 420 epochs

(C)(R) 20240622, R Gerthsson