This. These are largely mixing decisions, although it would be kinda weird to have a whispering actor boosted to sound as loud as a train in a scene.

The technical term for this is “Dialogue LRA” measured in LU: https://s3.amazonaws.com/izotopedownloads/docs/insight200/en... - this link also mentions some of the standards around this and now to use their software to adhere to those standards.

LUFS is also relevant here as a measure of loudness: https://blog.landr.com/lufs-loudness-metering/

https://auphonic.com/blog/2020/10/08/dialog-loudness-normali...

and

https://www.pro-tools-expert.com/home-page/2018/11/10/is-dol...

Give some insight into how this works.