Recordings come back with who said what
One call turns a recording into a transcript with each person's words. On a real four-person meeting, 96.7% of words went to the right speaker.
Transcribing a recording now places every word with a voice, rather than cutting at pauses and hoping each stretch is one person. A 22 MB speaker model does this for recordings and LS-EEND keeps doing it live.
We measured it on a public 17-minute meeting with four people: 96.7% of 1,912 scored words went to the right speaker and all four were found, in 13 seconds. The model we started from reached 86.3% and found three.