Data mining - How create a decision threshold

[mikiblu81]

Hello everyone, I'm making a speaker recognition system: I have a model gmm for each speaker; in the test phase I compare each model with the incoming voice and I get the scores for each comparison, eg.:

gmm_speaker_1 vs voce_input = -33.1
gmm_speaker_2 vs voce_input = -31.2
gmm_speaker_3 vs voce_input = -33.4
gmm_speaker_4 vs voce_input = -34.0

I always get a score, even if there is not a speaker that matches the voice. When the speaker is the voice input, the score is the lowest among all, in the example -31.2. I need to understand how I can create a statistical threshold or a range to decide whether the speaker is unknown, eg:

Unknown Speaker <-33
Probably Speaker > -33

the range of scores is -31 to -35

thank you very much

 

[Predictor]

Different thresholds will offer different trade-offs between errors and throughput: at one end, errors are less frequent but sometimes no outcome is chosen; at the other end, some outcome is always chosen, but errors are more frequent. You will need to decide what trade-off best suits your problem.

 

[mikiblu81]

Thank you very much for your reply. For the moment, my goal is to reduce the number of places and "errors are less frequent but sometimes no outcome is chosen," some idea about this?