Evaluation of methods to combine different speech recognizers

Abstract

The paper deals with the problem of improving speech recognition by combining outputs of several different recognizers. We are presenting our results obtained by experimenting with different classification methods which are suitable to combine outputs of different speech recognizers. Methods which were evaluated are: k-Nearest neighbors (KNN), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Logistic Regression (LR) and maximum likelihood (ML). Results showed, that highest accuracy (98.16 %) was obtained when k-Nearest neighbors method was used with 15 nearest neighbors. In this case accuracy was increased by 7.78 % compared with best single recognizer result. In our experiments we tried to combine one native (Lithuanian language) and few foreign speech recognizers: Russian, English and two German recognizers. For the adaptation of foreign language speech recognizers we used text transcribing method which is based on formal rules. Our experiments proved, that recognition accuracy improves when few speech recognizers are combined.