dominoanty / SpeakerRecognition

Implementing speaker recognition using Python (GMM-UBM)
28 stars 14 forks source link

Setting up Treshold #5

Open predestination opened 5 years ago

predestination commented 5 years ago

If I'm giving an unknown voice, it is still returning a speaker. As you mentioned in other issue that we have to set treshold. What I did is , I'm taking the max value from confusion matrix, and then checking if the value is atleast 70% of the sum of the confusion matrix. However I'm not getting results, moreover for an unknown voice I'm getting a very good confusion matrix like [1467, 690] which I never expected. I though that GMM-UBM will solve my unknown voice issue, and moved from GMM to GMM-UBM. Can you suggest some tips or throw some light with implementation is possible on how to tackle this? and am I doing the correct way for treshold?

Example for treshold: confusion matrix : [1130, 450, 789] 1130/2369. Here 2369 is sum of the matrix.

Now the value of 1130/2369 is 48% which is less than 70% , so I reject it.