itertracks() doesn't work properly in 0.1.1

[Purfview]

itertracks() doesn't work properly in 0.1.1. problems:

1. Hyperparameters onset/offset doesn't work properly, for example in the output there are segments shorter than "onset".
2. Looks like segments in time skew to front.
3. Segments overlap.

Audio to test:
audio.mka_dump.zip

UPDATE:
I thought that onset/offset corresponds to "min_duration_on"/"min_duration_off".
onset/offset shouldn't be touched for pyannote/segmentation-3
"min_duration_on"/"min_duration_off" could be set in some very first version of pyannote-onnx , can you add them back?

[pengzhendong]

You could try sherpa-onnx with pyannote-audio.

[Purfview]

You could try sherpa-onnx with pyannote-audio.

I tried it but it doesn't have "min_duration_on"/"min_duration_off" implemented too, btw, it works very well, almost same output as original pyannote segmentation.
Another thing it's slow, ~same speed as pyannote's torch version, when your repo is ~4 times faster.

[pengzhendong]

OK, I will try to add "min_duration_on"/"min_duration_off".

This repo is faster because I use a large step = 5s (pyannote's is step = 1s, you could change it):

pyannote-onnx/pyannote_onnx/pyannote_onnx.py

Line 108 in bcd0a7f

step = self.duration // 2

[Purfview]

step = self.duration // 2

Tested step = self.duration // 10 it got slower only ~10%