Let's talk about "NAM-to-X"

In the last few months, we've seen an explosion of products from different companies that have benefited from the open-source nature of NAM, enabling them to make their products offer more for musicians. For many of these, users can point the product at their favorite NAM models (held in a .nam file) and use that model's tone with the product.

Arguably, these all "use NAM" in a broad sense. However, the software behind the scenes can vary widely, and that can impact the result.

I'm proposing to coin the phrase "NAM-to-X" as a general term for products that integrate with NAM by converting the model to be used by a proprietary algorithm that approximates the behavior of the source model.

The term "conversion" has become a popular shorthand for describing what happens, and I think that people typically get the right idea with it. However, I don't prefer to use it for a few reasons.

First, and most importantly, it doesn't answer the question: "Conversion to what?" Not all conversions are created equal. Musicians have been very interested in comparing the various products and technologies available; here, preciseness helps--and so does brevity. "Conversion" is brief, but not precise; and in the past I've been precise, but not brief:

This January at the NAMM show, Valeton shared the GP-5, a mini-format pedal they had been working on that included all of the staple ingredients to a guitar signal chain, but also included an integration allowing them to use .nam files for the pedal's amp effect block...My understanding is that Valeton is using the same SnapTone modeling approach used in the Sonicake Pocket Master, so I expect a similar experience using NAM on this to that.

-from this post.

I think that "NAM-to-SnapTone" has the best of both. Additionally, I think this gives folks space to make up their minds about SnapTone on its own merits (and similarly for other proprietary technologies). If someone makes a high-quality conversion, then that deserves to differentiate that product--it shouldn't be lost on people!

That brings me to my second reason: "conversion" doesn't necessarily imply lossiness. Here, I'm talking about "lossless" in the "pure" sense--There are multiple ways of coding up the calculations that NAM does--and some products use a different implementation from what I've coded. These are technically conversions, but ones that aren't really related to the issue at hand.

Third: Conversion doesn't necessarily imply that the result isn't a NAM. One could "convert" a WaveNet NAM into an LSTM NAM or a ConvNet NAM, a different-sized WaveNet...or I've been somewhat curious to learn how to implement some of the more "classical" model architectures for fun, which might enable NAM's open-source code to offer an out-of-the-box "conversion" functionality from NAM for builders. I think there's an argument for this being distinct. There are upsides for musicians like increased transparency about how the converted model works and what to expect from it. The other is that, as part of NAM's open-source code, it could centralize improvements to the modeling tech itself so that more products (and, therefore musicians) could benefit.

Conclusion

There's been confusion about how various products building on top of NAM have integrated into the open-source project. I hope that "NAM-to-X" gives folks a nice, precise, and succinct way to talk about it.

Feedback is very welcome. Feel free to discuss on the NAM group on Facebook, and you can also get in touch with me directly.