Wouldn't C# be the most reasonable option if so? Maybe they did the math and the .NET ecosystem is much more expensive than Java? Even ignoring MS cost, most 3rd party libraries on .NET usually require a license.
I don't have experience with machine vision but here's an example of a speech2text library that integrates whisper.cpp in an idiomatic way: https://github.com/sandrohanea/whisper.net
There are many good community projects with code quality way higher than your average enterprise SDK with layers upon layers of abstractions and allocations. Finding them is the same as with most other languages like Rust or TS.