Note: I am the creator and maintainer of ar5iv.
I think the idealized goal should be that the ar5iv HTML can be successfully used with all mainstream accessibility tools, including VoiceOver - which was mentioned in the original question. Maybe it is worth reporting the awkwardness of the verbosity of speaking "maths" to their team, with an example document attached?
Personally, I have successfully tested with two combinations of AT for proof-listening, one OS-specific and one cross-platform. But I have done this as a developer troubleshooting where to make the markup better, so I can not give you an unbiased assessment of reading quality.
If you have access to a Windows device, one could use NVDA with the MathCAT plug-in for mathematics, as described here:
https://addons.nvda-project.org/addons/MathCAT.en.html
On any device, you could use the recent versions of the read-aloud browser extension, which uses a different AT for math syntax called Speech rule engine (SRE). I have tried that with Chrome, read-aloud v2.5.0
https://chrome.google.com/webstore/detail/read-aloud-a-text-to-spee/hdhinadidafjejdhmfkjgnolgimiaplp
Both seem to provide nominally acceptable readouts, but on a close listen we still have some artificial outcomes - sometimes due to not providing more in-depth markup, sometimes due to the tools not being sufficiently context-aware.
I welcome reports for such problems in ar5iv, when used with any accessible technology. An example of a VoiceOver report we still have to address can be seen here:
https://github.com/dginev/ar5iv/issues/95