Incorrect metric transfer about pyannote-audio

[hbredin]

You are right about metric parameter not being passed to the parent class. I will fix this in the next release. Thanks for catching this... this might actually explain weird behaviors I noticed in the past!

To speed up training, I suggest you reduce the number of speakers per_fold to something smaller like per_fold=20. All possible triplets are tested so it means that the number of tested triplets is proportional to per_fold^3.

You may also want to reduce per_label to only a few samples per speaker: per_label=3. It needs to be greater than 2 to be able to build positive pairs.

[Ruslanmlnkv]

Thanks so much, I'll try this.

[hbredin]

This has just been fixed in develop branch.
Thanks again for catching this critical bug!