Speech recognition and models
Diarisation
Splitting a recording into segments by who is speaking, without necessarily knowing who those people are.
Diarisation is splitting an audio recording up by who is speaking. The result is “Speaker 1”, “Speaker 2”, “Speaker 3”, with a timestamp for every switch.
The question diarisation answers is who spoke when. That is a separate job from the speech recognition that writes down what was said. The two usually run in parallel and are stitched together at the end.
Not the same as knowing who people are
Diarisation groups voices, it does not identify them. The system knows that two segments are the same person, but not that the person is called Kari. You put the names on afterwards. Tying a voice to a known identity is speaker recognition, which is a different task.
Where it breaks down
- A shared microphone. Five people around one phone in a meeting room is the hardest situation there is, and at the same time the most common.
- Overlapping speech. When two people talk at once, the system has to choose, and often chooses wrong.
- Similar voices. Two people in the same age group with the same accent and intonation are regularly merged into one speaker.
- Short interjections. A one-second “yes” is often too little material to place.
The simplest thing that helps is for participants to say their name at the start of the recording. It gives both the system and you something to attach the voices to.