Speaker diarisation
Splitting a recording by who is speaking at each moment: Speaker 1, Speaker 2… without yet knowing their names.
Diarisation answers the question “who spoke when?”. A model analyses the characteristics of each voice across the recording and groups the stretches that seem to come from the same person. The result is anonymous labels such as “Speaker 1” and “Speaker 2”, which you can then replace with names.
It’s harder than it looks: in a room with one microphone every voice sounds alike, people interrupt each other and some turns last a second (“yes”, “OK”). The typical errors are splitting one person in two or merging two people into one.
Echo separates speakers on its own servers and then rejoins a person who was split by subtracting what every voice in the recording has in common (the room, the microphone). In our test with conversations between known speakers, that raised accuracy from 80.7% to 85.0% without mixing anyone up.
If something goes wrong, you can change who says each paragraph in the editor or use “Re-separate speakers” in the Speakers tab. Don’t confuse it with speaker recognition, which puts a name to a known voice.
