Unknown speaker?

[Dmytro Prylipko]

Hi,

I have a simple question: is it possible to build a bob.bio.spear setup that will either recognise a speaker among enrolled ones or produce the UNK output label.

The latter case corresponds to the situation when the test footprint is not similar to any known profile. Sure, there should be a possibility to set the UNK-decision threshold.

If it is possible, what is the most natural way to implement it?

Thank you

[Manuel Günther]

Dear Dmytro,

interestingly, I lately have discussed this topic with some colleagues. What you want to have is an Open Set recognition, i.e., when the target speaker is not enrolled in the database.

The result of our discussion was that we might have special score values, for example ``-inf`` to signal that the target speaker is unknown.

However, so far, an unknown label is not included into ``bob.bio.spear``, and no evaluation mechanism is provided. Anyways, you are very welcome to be the first to introduce an evaluation measure to the Open Set problem into ``bob.bio.spear`` (particularly, the ``bin/evaluate.py`` is provided in the package ``bob.bio.base``)