Keith Bloemer

Good arguments, I’m guessing you mean can I tell the difference between a large model and a slightly smaller model? Probably not for every case. The tricky thing about plugins is you have to generalize them to work well on the most hardware/OS/DAWs. Where things like Kemper and Quad Cortex have full control because they own the hardware too. If I start selling plugins these are things I definitely want to think about. But for now I just made something interesting that I enjoy using, and decided to open source it because it turns out there’s a lot of smart people that want to help improve it. Sometimes it’s hard to separate the software aspect from the music, but you bring up a good point that the music and sound should come first before fancy technology.

The short answer is, it depends. Sound is subjective, and to my ears, I can hear the difference if I lower the parameters for high gain targets. The less distorted the sound, the easier it is to train, so for clean or low distortion I might get away with lowering the parameters. But I want it to work as well as it can for any potential sound, so it’s my best guess based on limited testing on the hardware I have available. User feedback on what works and what doesn’t has helped a lot! With the training, you’re trying to minimize loss, which is the difference between the predicted sound and the actual sound, so I can put a number value on the accuracy of the models. I also look at plots of the waveforms to visually see how close the sound is.

For this plugin, yes it’s necessary. It only runs on profiles, even the five included tones. The profiling works by training an artificial intelligence model, which basically amounts to adjusting a bunch of number values (you can see the values if you open the .json files in a text editor). The values are adjusted during the profiling until they can make the input signal get as close as it can to your target signal. The resulting “trained” model is the loadable .json file. I expect that as this tech develops it will get faster and more accurate, but as with any modeling it is still an approximation. I could make the model size smaller, which would reduce cpu usage, but it would sound worse.

Hey! Developer here, thanks for all the feedback and models! High cpu usage is currently the trade-off for the flexibility of copying pedal/amps. There's alot of number crunching behind the scenes that allow for the same model structure to produce a wide range of tones (as opposed to optimizing for one specific sound). From what I've heard from the Quad Cortex, my SmartAmp plugins are not quite there in terms of accuracy to the target hardware, but hoping to improve it. Also, the SmartAmpPro is intended to be used with reverb or IR's, so if you're running it by itself it will sound dry. Here's a model I made from recordings of the Gojira plugin; which definitely gave the capture some difficulty. The result is a pretty good sounding distortion model, but not 1:1 with the Gojira sound. Enjoy! gojira_test_model.json