FLUERS and GP's Common Voice dataset focus on read speech. I've observed models that perform well on these datasets be completely useless on other distributions, like whispered speech or shouted speech or conversational speech between humans who aren't talking to a computer.