Likewise, the Common Voice English dataset isn't great for ASR training outside India, either. There's a huge proportion of Indian speakers, and their data doesn't really help train ASR systems for non-Indian accents.
You and OP ("the indians labelers don't know how to do it correctly") seem to want the former, so it would be good to state that goal upfront.
A model is more useful if it approximates the speaker’s intent, not the listener’s interpretation. “Right” is complicated in language, but it’s hard to see how you’d use a model full of cross-cultural misunderstandings.