dimensions of MLLR matrix

[Balaji]

Hello,

I am performing "Adapting the acoustic model" as explained in the tutorial.
I am using continuous model while executing bw command and map_adapt.exe.
The matrices created are of the following dimensions:
Means, Variances, Mixture weights, Gaussian densities: are all 5138 X 1 X 32
As I understand, 5138 is number of senones. What is this 32? The feature vector's dimension is 13 + 13 +13 = 39 right?

Also, the MLLR Matrix is 36 X 36. How?

Can somebody explain please? Thank you.

Balaji.

[Nickolay V. Shmyrev]

As I understand, 5138 is number of senones. What is this 32?

32 is number of gaussians per senone. Each GMM model is a mixture of gaussians.

The feature vector's dimension is 13 + 13 +13 = 39 right?

Yes

Also, the MLLR Matrix is 36 X 36. How?

Before processing features are multiplied by LDA matrix feature_transform (39x36), so the dimension becomes 36.

[Balaji]

Thank you.

[Balaji]

Hello Nickolay,

In this thread, earlier you wrote:

32 is number of gaussians per senone. Each GMM model is a mixture of gaussians.

Any justification for the number of gaussians being 32? Can it be more or less?

Before processing features are multiplied by LDA matrix feature_transform (39x36), so the dimension becomes 36.

Similarly, in feature transorm - the number of features (feature vector's size) is 39. But, how is this 36 chosen.

Thank you.