Text this: Benchmarking the First Realistic Dataset for Speech Separation