Issue link: https://digital.copcomm.com/i/933783
Neural Networks: Their magic depends on a new concept for pro audio: neural networks. This technique has exploded, driven by demands of speech recognition and image processing in consumer applications. As studio computers get more powerful, and software developers get more experience with this kind of computing, we'll see a lot more applications in production and post. Neural networks were proposed as theory in the 1940s, and the first simple ones appeared in labs in the 1950s and '60s. But they were limited by the computers of the time and not really practical. Studio engineers had to rely on standalone analog filters, compressors, and other hardware, patching them into channels of the era's limited consoles. If we 2 wanted more complex effects, we'd string multiple processors in series (such as putting a tape delay before a reverb, to simulate the slap of a stadium PA). We've come a long way since analog mixing, but we still "route signals through processors." Our DAWs and digital consoles let us drop software plugins into a channel. Or we route clips through plugins offline, replacing each clip with a processed version. Most of those plugins were designed using a similar process! Developers create a signal path made of basic functions—envelope detectors, delays, mixers, and so on—feeding one another. This becomes the processing recipe or algorithm. They build it by stringing virtual modules together in a programming environment. Figure 1a shows how this worked for a simple plugin that simulates a 16mm classroom projector—complete with audio degradation, wow and flutter, gate chatter, and a motor that ramps up to speed when you turn it on. I wrote the processor in the open-source SonicBirth application a few years ago, when I had to futz a bunch of elements for a project, and the figure is a screenshot from that program. I added the red labels so you could trace the signal flow more easily. The rectangles with just one wire are user controls or numeric constants; most of the other rectangles are single processing functions like generators or filters, but some are shortcuts to other multi-processor algorithms (like the envelope follower in the top row). Figure 1b shows the plugin's control panel; despite all the internal connections, the user sees an integrated whole. This is a simple plugin, thrown together to solve an immediate need. Sophisticated commercial processors can have hundreds of internal modules. T wo new sets of audio tools promise things that were previously impossible, like automatically isolating—rather than just suppressing—non-dialogue elements in production recordings or mixed tracks. The good news? Audionamix's TRAX Pro 3, XTRAX Stems, and SVC, along with iZotope's RX 6 Dialog Isolate and De-rustle modules, actually do what they promise 1 ! There are limits, of course: while these programs give you unprecedented control in ideal situations, real-world challenges may produce unpleasant artifacts if you turn them up too high … just like almost any other tool in the studio. by Jay Rose CAS A NEW WAY TO THINK ABOUT PROCESSING Wow generator Distortion Flutter generator Bandpass Speed modulation Motor Ramping 24 fps clicking gate noise Figure 1a: The algorithm for a simple processor, expressed as a circuit diagram in SonicBirth. 1 www.audionamix.com and www.izotope.com. While this article is about the technology, some details of these programs appear on the last page. Re-recording mixer Stephen Fitzmaurice CAS has contributed his first impressions of the Audionamix software on page 35. And we've posted before/after samples from both systems at http://cinemaaudiosociety.org/over-the-net/. But these aren't comprehensive reviews … those will be coming in future issues of the Quarterly. 2 Yes, "we": my first studio gig was in the 1960s. I'm that old. Figure 1b: That processor's control panel.