I tried to understand whether CAKE is the right way when it comes to the specific requirements of prioritizing VoIP and streaming over anything else, and the following contrast made me laugh:
https://www.bufferbloat.net/projects/codel/wiki/CakeTechnical says: Cake’s fundamental design goal is perceived simplicity.
https://help.mikrotik.com/docs/display/ROS/Queue+types says: CAKE (Common Applications Kept Enhanced) … Disadvantages: More complex to configure than simpler queuing disciplines. Not ideal for all use cases, particularly where fine-tuned control is required.
On a serious note, what I’ve understood from the manual is that if using CAKE, you don’t need to (or rather must not) deal with one leaf queue (selected using a corresponding packet mark) for each priority class; instead, you set the DSCP headers to tell the single common CAKE queue which “tin” to put the packet to.
This post seems to have it all, except the way to handle the WAN->LAN traffic independently for each source WAN; there, the packet marking remains necessary in addition to the DSCP marking. What makes me more cautious is the fact that in the OP’s requirements outline multiple LAN interfaces but the WAN download bandwidth is (naturally) common for all of them, which does not play well with the need to use the interface queues as the parent ones. So I wonder whether it would make sense to use a hairpin tunnel to overcome this conflict of requirements.