Multispeaker banks work by adding a speaker id to the end of the line of the transcription. The speaker ids separate the different voices. The current pipeline notebook does this automatically as you load different datasets for your multiple speakers.