SpeakerNet Speaker Identification: how to identify unknown speaker?

[JINHXu]

Hi!

I am using the SpeakerNet model for Speaker Identification, it appears that in the inference stage, the model always returns a label, however, I was hoping to also identify that "a stranger" is "a stranger", i.e. for a speaker that is not in the training data, I expect the model to also realize that the speaker is unknown. But in this case, the model always returns the most similar label instead of unknown. I have a few thoughts myself (mentioned here), and I was wondering how is everyone else handling the situation? Ideas would be much appreciated!

Cheers!
Xu

[JINHXu]

To answer my own question, this can be done by adding an Unk class of various random speakers (no overlap with other speakers in the training data). I went with recordings from voxceleb, 3-4 audio samples from each speaker, the results look good on recognizing unknown speakers. In case this is helpful to anyone else.

[tkcoding]

@JINHXu sorry for the late comment here.
I have the thought of predicting the speech with probability output (logits) and from there understand the threshold instead of getting max value of the probability. For e.g. if a similar voice typically the max probability will be > 0.7 if model produce max probability < 0.5 then we can put this into "stranger" class.