| 摘要 |
The invention provides a method and apparatus for adapting a pattern recognition model, specifically a speech recognition model, between first and second environments. A model for performing pattern recognition on an inputted sequence of observations is initially provided, the model having been trained to recognise a pattern in a second clean noise environment. The model has a plurality of parameters relating to the probability distribution of a component of a pattern being related to an observation. The model is then adapted S59 to a first noise environment using inputted observations S51 from the first noise environment. Adapting the model trained in the second environment to that of the first environment comprises using second order or higher Taylor expansion coefficients derived for a group of probability distributions an wherein the same expansion is used for the whole group. The groups may be regression classes. In order to recognise speech a language model is also used with the combined likelihoods of the adapted model and language model used to output a sequence of words identified from the input signal of the first noise environment.
|
