Most people working with speech AI don't build their own datasets. They pick one and find out later whether it was the right choice, usually when the model fails in production in ways the benchmark ...