I would strongly consider Nextcloud AIO instead of building and maintaining the Talk HPB stack separately.
AIO already includes Talk HPB/TURN as part of the managed stack, so there is no need to automatically introduce another server just for HPB.
For a normal deployment without call recording, a modern 8-core CPU and around 16 GB RAM is already a very reasonable starting point. Obviously, an 8-core modern Ryzen/Xeon is a very different thing from an old 8-core server CPU, so the CPU generation and actual performance matter as well.
The important question is whether you need Talk call recording. Recording is much more CPU-intensive because the recording server has to encode the video during the call. If recording is required, I would size the hardware more generously and benchmark it for the expected number of simultaneous recordings.
Nextcloud’s own recording documentation confirms that video encoding is typically the main CPU load.
I have also done some practical AIO testing here:
AIO: