2dor
Roadie
- Messages
- 897
That xSTD architecture is cool if it's the one labeled MY8886_16 in his core.py file.10% upon the standard architecture. If standard is 100%, xSTD is 110%
EDIT: it's actually a bit less than 10%
View attachment 37391
version_12 = reamp trained with the xSTD (MY8886_16) architecture
version_13 = same reamp trained with the vanilla Standard architecture
A couple of conclusions I can draw:
- training takes a bit longer for 1000 epochs but not crazy much (~8 minutes on my RTX 4070, AMD 5700X setup)
- high-end seems to be more faithful to the source (lower pre-emphasized MRSTFT) which is pretty cool
- ESR shows improvements too which is nice as well; shows that most of the spectrum (aside from the high-end) is faithfully represented
- The Standard vanilla architecture shows about 14% - 15% RT-CPU in Reaper on my machine
- The xSTD architecture shows a slight bump to 17% - 18% RT-CPU in Reaper on my machine
Overall, I think this new architecture combined with the enhanced "anti aliasing" sequences can help models behave better as far as the overall training is concerned (high-end and overall time-domain).
Last edited: