Analysis of Disturbed Acoustic Features

An analysis method was developed to study the impact of training-test mismatch due to the presence of additive noise. The contributions of individual observation vector components to the emission cost are determined in the matched and mismatched condition and histograms are computed for these contributions in each condition. Subsequently, a measure of mismatch is defined based on differences between the two histograms. By means of two illustrative experiments it is shown to what extent this emission cost mismatch measure can be used to identify the features that cause the most important mismatch and how in certain cases this type of information may be helpful to increase recognition accuracy by applying acoustic backing-off to selected features only. Some limitations of the approach are also discussed