User:Icax0r/Maximum likelihood linear regression

Maximum Likelihood Linear Regression (also known as MLLR) is a commonly-used technique in Speech recognition for speaker adaptation. Given a small amount of speech from a target speaker, the parameters of a speaker-independent model can be adjusted to maximize the likelihood of the target speaker's speech under the model.

Description

edit

Parameter Tying

edit

If not all of the distributions are observed in the training data, then the parameters may be tied - i.e. a single transformation matrix is applied to several distributions. Distributions can be clustered according to phonetic similarity, or automatic clustering methods may be used.

-- TODO -- math

Relation to Other Types of Adaptation

edit

CMLLR

edit
-- TODO -- doesn't rely on having a transcript.  and MLLR does??  use a GMM to estimate the tr

MAP Adaptation

edit
-- TODO you need more data. --

References

edit
edit