Acoustic model Training-lda algorithm __lda

In speech recognition, in order to enhance the robustness of audio features, it is necessary to extract the characteristic vectors with strong distinguishing ability, and the common method is PCA and LDA algorithm.

The PCA algorithm seeks to preserve the most effective and important component of the data, leaving out some redundant components that contain less information.

LDA is a change matrix to achieve the goal of dimensionality reduction, the difference between LDA and PCA is that LDA allows the distribution of the sample to condense, so that the distribution of samples between the alienation, so that the characteristics of more representative, in the classification of the LDA algorithm better.

As shown in the following illustration:

LDA has the following assumptions:

1 the feature vector projection not all dimensions contain the information of distinguishing force, they are included in the former P Wizi space, and then (N-P) Wizi space does not contain useful information and is ignored;

2 Each sample is a Gaussian distribution.

The formula is defined as follows:

When speech features are represented by vectors, the LDA algorithm hopes that the covariance matrix between the model samples will be greater and better,

The covariance matrix in the sample is as small as possible, so the wopt takes the maximum value.

The column vectors of a projection matrix are: The eigenvectors of the D maximum eigenvalues (eigenvectors of matrices)

The covariance matrix after LDA transformation cannot be diagonally, so it needs to be transformed by Mllt (Maximumlikelihood Linear Transformation, mllt).

An introduction to the MLLT algorithm is presented in detail in the next article.