Change number of Gaussian component when enroll speaker using MAP adaptation

[Nguyễn Cẩn]

I trained UBM with:

        ubm_model = algorithm.GMM(2048)

        ubm_model.train_projector(my_features,'ubm_model.hdf')

Follow this MIT document, they said UBM is a large GMM (2048 mixtures) trained to represent the speaker-independent distribution of features. Hence, I used 2048 Gaussian component to train UBM.

Then, I enroll new speaker using MAP adaption: 

       speaker_model  = ubm_model.enroll(speaker_features)

At this step, I'm wondering that how many number of Gaussian component I have to use. I look at GMM source and I found that bob using number of component of UBM to enroll new speaker. So, it gonna use 2048 Gaussian components to enroll.

I don't know it's right thing or not to enroll new speaker with 2048 Gaussian component. I read some papers, they just use 64 or 128.

Can anyone help me to solve this question? Can I change with lower Gaussian component with bob.bio.gmm when enroll new speaker?

Thanks,

[Manuel Günther]

Dear Nguyen,

I do not know, where you got the numbers 64 or 128 from, they do not appear in the document that you linked. These numbers are theoretically not possible -- at least not when you use MAP adaptation from the UBM to enroll a model. With MAP adaptation, you will always enroll a model with 2048 Gaussian components -- given that your UBM contains this many Gaussians.

This should answer your question: No, you cannot change the number of Gaussians in an enrolled model.

Note that the number of 2048 Gaussians inside your UBM should only be that high if you have enough training data. When your training set is small, such a large number of Gaussians might tend to overfit to your training set, so any enrollment will be poor.

Best regards

Manuel

[Nguyễn Cẩn]

Dear Manuel Günther,

Thanks for your quick response. It's very helpful.

Best,

Can Nguyen