the simpliest way to “port” existing NAM models to Aida is to use the provided input.wav/output.wav files and re-train using our training script. Which is not dissimilar to the training script used by NAM, since in the end we both use pytorch and Jupyter notebook to drive the training process. We went a bit further maybe since we now provide the possibility to do training in a docker container, which is really convenient imho. We will provide instructions on “how to retrain NAM models for Aida DSP”. The quality however, at least from “numbers” perspective, would probably decrease when re-training with the network used with Aida DSP. The problem is that if the NAM dataset works really well with their Wavenet network, the same dataset with another network type (In Aida we use RNNs) could provide different results. Infact, we use a completely different dataset to train our models.
the harder way would need probably to discuss things with original NAM developers. But the basic idea would be to “merge” the training scripts. And more important to let them export the model in a format that is compatible with RTNeural. As developers, we don’t want to have multiple/custom inference engines implemented in every neural plugin. We want to converge toward a single and flexible inference engine with a focus on rt-audio (which is, imho, RTNeural) and possibly have everybody to contribute to the quality of this core piece of every neural plugin. Or at least this is my proposal. However even considering a script to convert NAM model to Aida DSP or whatever, the main problem remains that NAM models are Wavenet models which are indeed too heavy for embedded systems. This is a known problem also for NAM developers, but currently they’re focusing more on VST kind of users.