Spoken Language Identification with Prosodic Features

This thesis focuses on the use of prosodic features for automatic spoken language identification (LID). LID is the problem of automatically determining the language of spoken utterances. After three decades of research, the state-of-the-art LID systems seem to give a saturating performance. To meet the tight requirements on accuracy, prosody is proposed as alternative features to provide complementary information to LID.;There are no conventional ways to model prosody. We use a large prosodic feature set which covers fundamental frequency (FO), duration and intensity. It also considers various extraction and normalization methods of each type of features. In terms of modeling, the vector space modeling approach is adopted. We introduce a framework called prosodic attribute model (PAM) to model the acoustic correlates of prosodic events in a flexible manner. Feature selection and preliminary LID tests are carried out to derive a preferred term-document matrix construction for modeling.;The PAM-based prosodic LID system is compared with other prosodic LID systems with a task of pairwise language identification. The advantages of comprehensive modeling of prosodic features is clearly demonstrated. Analysis reveals the confusion patterns among target languages, as well as the feature-language relationship. The PAM-based prosodic LID system is combined with a state-of-the-art phonotactic system by score-level fusion. Complementary effects are demonstrated between the two different features in the LID problem. An additional operation on score calibration, which further improves the LID system performance, is also introduced.