How well does this work with conversational speech? Any benchmarks?