Do we need to run VAD before running pyannote - speaker diarization pipeline ?

[amitli1]

According to pyannote.audio: neural building blocks for speaker diarization (pyannote version 1) the pipeline includes VAD blocks.

According to pyannote version 2.1:
PYANNOTE.AUDIO 2.1 SPEAKER DIARIZATION PIPELINE: PRINCIPLE, BENCHMARK, AND RECIPE

they overhauled the speaker diarization pipeline and removed some blocks (including VAD, overlapped speech detection).

So if I'm using pyannote - speaker diarization pipeline (version 2.1.1),
Do I need to run VAD and cut the speaking segments before running the speaker diarization pipeline ?

[hbredin]

No, you do not need to run VAD before.

Also, even if you do have your own VAD, there is no easy way to pass it to the pipeline.

[amitli1]

So segments without speech will be concentrated to one cluster ("no speech cluster") ?

[hbredin]

No. There will simply be gaps in the output for non-speech regions.