Task analysis methods for data selection in task adaptation on mandarin isolated word recognition

This paper studies the performance of different task analysis methods for data selection in task adaptation on Mandarin isolated word recognition. Two key issues are investigated including 1) the type of the coverage units; 2) the method to generate the distribution of the acoustic units (coverage units) by task analysis. For the first issue, three coverage units namely word, syllable and right-context-initial/toned-final (RCI/TF) are used and compared which cover different context information. For the second issue, the traditional coverage unit balanced task analysis approach is compared with a new developed automatic analysis of vocabulary confusability. In our experiments on Mandarin isolated word speech, it is observed that performance with task analysis is much better than without task analysis. Comparing with RCI/TF and syllable, the word performs better as the coverage unit due to more context information involved. Moreover the active approach performs worse than traditional coverage unit balanced approach which means more accurate approach for the confusability analysis is necessary.