I think what you have observed is expected. Even though you’re using a lower sample rate on mobile, FADPCM trades compression efficiency for lower CPU usage, which often results in larger bank sizes compared to Vorbis.
You are not doing anything wrong, I have found some related discussion that might help you dive deeper into optimization strategies:
Hope this helps, let me know if you have questions.
Thank you for answering.
This makes more clear for me about the bank sizes then. I’ll have to dig deeper into the encoding formats and find the best memory/cpu balance.