Accentize dxSplit

We check out the latest of Accentize’s excellent dialogue restoration tools.

Accentize’s focus is on developing plug‑ins that apply machine learning techniques to the problem of noisy dialogue. And by noise, I mean anything that might compromise the sound, including actual noise, reverb, or changes in tonality due to filtering or other processes, such as when recording a phone call. Only founded in 2019, they’ve already won multiple awards for their plug‑ins, including an Emmy for dxRevive — a wonderful tool that I reviewed in SOS March 2024. That plug‑in has since been updated, and is directly relevant to this review too...

Split: The Difference!

At first glance, dxSplit resembles vocal isolation tools from Acon Digital (Extract:Dialogue 2) and iZotope (RX’s Dialogue Isolate module). While the concept is, indeed, broadly similar, there are differences. All these plug‑ins split the incoming audio into three, and you’re given separate fader control over the level of the voice, the related reverb/ambience, and the rest (noise). In dxSplit, those fader adjustments can range from +12 to ‑36 dB. There’s a clear scale alongside, but you can also enter the value directly in a box at the bottom. Below each fader are mute and solo buttons, and chain icons/buttons between channels let you link them so that, for example, moving the Voice or Reverb fader moves both, but Noise remains unchanged. At the bottom of the GUI is a four‑band parametric EQ for each stream, but there’s no spectrum sensitivity adjustment as on Extract:Dialogue 2. More on that comparison later.

It’s possible to load Accentize’s dxRevive plug‑in in within the dxSplit GUI, making it easier to use the two in combination.

Something unique to dxSplit is that if you have dxRevive, you can load it within dxSplit — it appears on the right of the GUI. Don’t dxSplit and dxRevive do a similar thing? Yes, similar, but not the same: dxSplit filters the source to derive the three components and adds nothing; dxRevive uses resynthesis, allowing it to generate new information to restore/fill out missing low‑ and high‑frequency information. And while the GUIs of and techniques behind dxSplit and its competitors are similar, obviously they’re trained on different data sets, so the results aren’t quite the same.

I have to say that dxSplit does a remarkable job of sorting the ‘wheat’ from the ‘chaff’!

Speak Easy?

I’ve used dxSplit and the other two plug‑ins mentioned above extensively over the last month or so, and I have to say that dxSplit does a remarkably good job of sorting the ‘wheat’ from the ‘chaff’. It might not offer so much user adjustment as Acon’s tool, but that’s kind of the point: it does such a stellar job of the separation in the first place that I rarely wish I could move content from one stream to another, and the rebalancing always sounds smooth and natural: it just works! And that’s the default Core Mode. There’s also an Extended mode, which delivers higher‑quality offline separation, using a Mac’s Silicon Neural Engine when you’re rendering. Occasionally, on some particularly challenging dialogue that was ‘self recorded’ by inexperienced remote podcast participants, I struggled to achieve the improvement I desired. But that I could achieve anything like a ‘result’ was remarkable, and running dxSplit and Acon’s Extract:Dialogue 2 in series allowed me to make a poor recording sufficiently intelligible that I could use it. dxRevive is a wonderful tool too, and the combination of that and dxSplit really is immensely powerful. It feels like cheating!

There’s enough latency involved that dxSplit and dxRevive are very much post‑production tools. But what tools they are! Turn the clock back just five years, and you’d be astonished at the progress that’s been made in dialogue processing. And in making such effective tools that are also so quick and intuitive to use, Accentize are among the ones really driving that progress. It doesn’t come cheap, but if you work professionally with dialogue, it’ll quickly pay its way.