I really like your strategy for edges. I think memorizing the first eight letters with audio and the rest with words (that critically, aren't necessarily one syllable, so that I wouldn't have to make a new list), would work very well for me. Or maybe I could do the first 4 targets with words and the rest with audio, I'll have to try them both out.I also have had this problem in the past, Here's what i do now.
I use a P.A.O system for corners (Can chuck locations in also if it is relevant and strengthens the image) and for edges i always do the first 8 with audio and the last ones (If any) with two fast words that don't need to make much sense, like DGOJ would be DoG OJ (Orange Juice or OJ Simpson). I also memo corner twists as a specific person for each corner and they are displayed a certain way depending on rotation direction, Edge flips are purely visual (I leave my fingers on the flips during the memo and just before dropping the blindfold i "Snapshot" where my fingers were).
Before i started doing this, about 2 months ago, i would use sentence memo for corners and full audio for edges with flips and twists being visual. My memo splits (according to the times i was getting just before switching) were 10-20 seconds for corners and 20-35 seconds for edges with somewhat poor accuracy, unless i was aiming for accuracy but then i would be going into the higher memo time.
Looking at today's session data i am getting 7 seconds for corners and 9-13 seconds for edges with some going over these and some going as low as sub-15 memo. But that seems to be the general area. Accuracy also increased quite dramatically a couple of days after switching but that was also the same day i decided to switch parity methods.
The best thing i like about this is that when i'm doing the first 8 edges i can rehearse the corner memo quite easily at the same time and then when i am doing the last of the edges i can do a quick recall of the first 8 and then finally snapshot any flipped edges.
I don't know if this ramble helped at all but this is the way i have found that works best for me. P.A.O images for corners mean i can go as fast as i like and not worry about strengthening, but it did take about 2 months to make a full list and their's still some pairs that are not fluent.
Good luck, man!
This does mean that I need to add more letters to my letter pair list, in order to accommodate the different buffer, which might be problematic. For example, for "X" I am often using "A" when I can't find an image that uses "X", which is fine since my buffer has "A". However, if I implement this, I would have to make sure not to confuse "A" with "X" pretending to be "A", if that makes sense. I mean this is something I'd have to do anyway if I wanted to get into bigger cubes BLD, which I think I will. (quick sidenote: I plan on using a UB buffer for when I switch to 3-style, which has my "A" letter, so this is a short term problem)
Thanks for ramble and for the luck!