Speaker adaptation for confusable words

摘要

A well-known problem in state-of-the-art speech recognition system is that often pairs of words occur that are very similar to each other and thus are confusable. This may cause errors in the recognition phase and thus decrease recognition rates. If unsupervised speaker adaptation is used in such a system, these misrecognitions may cause adaptation of the wrong models and thus cause a further decrease in performance. Therefore, according to the present invention, confusable words within the vocabulary are marked and an adaptation of the system to a certain user with such marked words is only performed in case of a positive confirmation of the recognition result by the user.