The mismatch between the model and operation conditions always exists. One of the main mismatch factors is speaker variability. As an example, Speaker-Dependent (SD) systems always provide better accuracy than Speaker-Independent (SI) systems when a large amount of SD training data is available. Speaker adaptive techniques can diminish the gap between these two configurations with a small fraction of the speaker-specific adaptation data. Maximum Likelihood Linear Regression (MLLR) method, one of these techniques, has been widely used to obtain adapted models for a new speaker when the adaptation data is not sufficient. However even it cannot completely remove the mismatch of speaker variability. For more reducing the speaker variability, it is more efficient to use the features enhancing speaker-dependency. These can be obtained by feature transformation based on Independent Component Analysis (ICA).
Since SD feature transformation matrix may be biased when adaptation data are limited, it is not always reliable. Thus we need to smooth the ICA-based feature transformation matrix applying to both the adaptation and test data. As a smoothing method we proposed to use a linear interpolation between SI feature transformation matrix and SD feature transformation matrix. From experiment results, we observed that the proposed technique is effective in speaker adaptation.