Doremi note: split stems first, then mash up for video VO

The memo frames its advice around a common sticking point: wanting one song's vocal delivery but another song's beat for a given piece of footage. It says that when full mixes are layered directly, both tracks assert themselves and collide with the voiceover, which is why separation is placed ahead of any mashup step.

It also ties the practical UI question to the same root issue, noting that source separation has long been worked on in both research and implementation, with Deezer's open Spleeter cited as a practical starting point. For product decisions, the memo says the test is whether a part you want to mute can be handled on its own.

The operational cautions are modest: keep one intent per pass so you can tell what worked, cap candidates at two or three, and check solo and mute in a web player while listening under the voiceover, since headphone solo checks can misjudge density. The stated goal is not polished music production but converging on one review track whose length and density fit, without full-mix collisions.