low-power-speech-recognition VS tinyusb

Some boilerplate but it's not too bad. Example: https://github.com/fb39ca4/low-power-speech-recognition/blob/master/scripts/toolchain.cmake

Implementing the required USB host interface is another hurdle, but tinyusb seems to have sorted most of that out, except that UAC2 is implemented as a device while you need a host.

Have a look at TinyUSB. It supports the STM32WB and demo programs for the Nucleo-WB55RG. It even offers an example for dual CDC ports, so very close to what you're looking for.

Second: Distrust, misunderstanding, and out of date opinions of C++ in embedded spaces. I often see "compatibilty" thrown around as a goal for using C and not C++, but for most modern embedded systems, C++ compilers exist and are (usually) well tuned. Arm is a shining example here*. C++ is more optimizable than C as the intent can be clearer to the compiler. C++ constexpr and templates are a godsend for embedded systems, and I used both heavily in my recent arduino and pico testing. I combined both of them for a really cool experimental USB interface descriptor builder that runs at compile time and avoids the need to count byte sizes, but I gave up on it when the maintainer said C only, despite the fact that C can't do that. I was similarly disappointed when the pico "C++" was C only, as there was no backwards compatibility necessary.

I looked at this but the fork seems to have had a lot of problems from looking at the PR: https://github.com/hathach/tinyusb/pull/1219

I think a better solution might be to use a USB flash drive in the usbc port. Unfortunately micropython and circuitpython don't support this (yet), so you'd have to use c++ https://github.com/hathach/tinyusb

It's possible to make the s3 be a mass storage device via TinyUSB. https://github.com/hathach/tinyusb/tree/master/examples/device/cdc_msc_freertos