Spoken Language Identification With Hierarchical Topic Model and Multi-levels Discriminative Cues

The discriminative cues and effective representations are two important factors for spoken language identification (LID). In this paper, the multi-levels language cues, including phonetic, phonotactic and prosodic, are studied, and a hierarchical representation based on supervised topic model is further designed to organize the multi-levels cues. The difference of languages is distinguished by using the distribution of prosodic topics and phonemic topics in each language. We built the proposed LID method and report the test results on NIST LRE07 datasets. Compared with the most advanced LID methods, our proposed method shows the competitive performance in the LID task. In particular, our method helps to capture robust discriminative information for short duration language identification.