Prediction of American listener's misrecognition of English words spoken by Japanese

摘要

This study tries to automatically estimate the probability of individual spoken words of Japanese English (JE) being perceived correctly by American listeners and to clarify what kind of combinations of segmental, prosodic, and/or linguistic errors are more fatal to the correct recognition. Firstly, from a large speech database of JE, a balanced set of 360 utterances of 90 male speakers were selected. Secondly, a listening experiment was done where 6 Americans were asked to transcribe these 360 JE utterances. Next, using speech and language technology, values of many segmental, prosodic, and linguistic attributes of words, which are related to the pronunciation errors, were automatically calculated from the JE utterances. Finally, relation between the misrecognized words and the attribute values were analyzed with Classification And Regression Tree (CART) method to automatically predict the probability of each of the JE words being correctly transcribed. The prediction performance was compared with the human prediction performance which was obtained by another experiment, where a Japanese teacher of English was requested to estimate the probability by hearing the utterances while looking at the intended sentences.