Andrew Rosenberg, Queens College, New York
Julia Hirschberg, Columbia University, New York
Native-like production of intonational prominence is important for spoken language competency. Non-native speakers may have trouble producing prosodic variation in a second language (L2) and thus, problems in being understood. By identifying common sources of production error, we will be able to aid in the instruction of L2 speakers. In this paper we present results of a production study designed to test the ability of Mandarin L1 speakers to produce prominence in English. Our results show that there are some consistent differences between the L1 and L2 speakers in the use of pitch to indicate prominence, as well as in the accenting of phrase-initial tokens. We also find that we can automatically detect prominence on Mandarin L1 English with 87.23\\% and an f-measure of 0.866 if we train a classifier with annotated Mandarin L1 English data. Models trained on native English speech can detect prominence in Mandarin L1 English with an accuracy of 74.77\\% and f-measure of 0.824.