search

iiscleap / NeuralPlda

Implementation of Neural PLDA (NPLDA) model (A discriminative backend for Speaker Verification)
98 stars 27 forks source link

questions about nplda #5

Closed forwiat closed 3 years ago

forwiat commented 3 years ago

hello iiscleap, Thank you so much for the great work you've shared. I have a question about the model loss. firstly, can I simplify the model loss to 'function(sigmoid(scores), target)'? (I don't konw if i understand it ...). Then, the scores can be regarded as loglikelihood radio? If can, how to explain it?

Thanks you!

iiscleap commented 3 years ago

Hi, Thank you for showing interest in our work. The loss function that we use is a soft version of the detection cost function (DCF), which we denote as softCdet. The threshold at which this loss is computed is one of the model parameters. Hence it is arbitrary. In this case, when you write this function in the form 'function(sigmoid(scores), target)', then the scores can not be treated as a log-likelihood ratio. The scores can be treated as log-likelihood ratios only for certain functions which penalize for bad calibration as discussed in the paper http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.217.7404&rep=rep1&type=pdf. An example of such a loss function is the Cllr. Regards.

forwiat commented 3 years ago

Oh! Thanks a lot! this paper may be very useful for me. ( BTW, the problem that bothers me may be solved, haha:)