How to add a new speaker without retraining

[IsmaelAh]

hello,i would to know when a new speaker (not in data) ,how can deal with this state ,i read about this in paper (Speaker Verification Using Adapted Gaussian Mixture Models) to get final UBM combie Individual subpopulation models ,i mean that is ,instead of add audio files to data(for a new speaker) then train from zero model, instead of that in paper they say:combine first UBM(before add new file to data) and combine it with second (new speaker) ,is this true or false,and if is true how can i do this using sidekit or other method,becuse not logical to train model from zero.

[Anwarvic]

That depends on what do you mean by adding a "new speaker".

• If you mean to train our model to identify the new speaker, then this speaker must be included in the training data and retrain the model from zero again.

• If you mean to verify the existence of this "new speaker", then that's the purpose of this repo in the first place and you can do that easily by scoring the likelihood of this new speaker. If the model is well-trained and working properly, you should get a very low likelihood (below 0).

[Anwarvic]

And if you want a system that can learn to identify the new speaker without being trained on him/her, then you should use something like Siamese Network. This network architecture is able to know the similarity and differences between the two input data. This network is widely used for face recognition but can be used for voice recognition as well.