For the evaluation in the paper, speakers are on separate channels (mono, it's a telephone conversation, after all). Generally there are solutions for separatin

For the evaluation in the paper, speakers are on separate channels (mono, it's a telephone conversation, after all). Generally there are solutions for separating speakers on a single channel that can work fairly well (assuming your training data is similar to the target domain) if you know the number of speakers beforehand, but it's tremendously hard if you don't (think transcription of large meetings).