EUROSPEECH 2003 - INTERSPEECH 2003
8th European Conference on Speech Communication and Technology

Geneva, Switzerland
September 1-4, 2003

        

Use of Trajectory Models for Automatic Accent Classification

Pongtep Angkititrakul, John H.L. Hansen

University of Colorado at Boulder, USA

This paper describes a proposed automatic language accent identification system based on phoneme class trajectory models. Our focus is to preserve discriminant information of the spectral evolution that belong to each accent. Here, we describe two classification schemes based on stochastic trajectory models; supervised and unsupervised classification. For supervised classification, we assume text of spoken words are known and integrate this into the classification scheme. Unsupervised classification uses a Multi-Trajectory Template, which represents the global temporal evolution of each accent. No prior text knowledge of the input speech is required for the unsupervised scheme. We also conduct human-perceptual accent classification experiments for comparison automatic system performance. The experiments are conducted on 3 foreign accents (Chinese, Thai, and Turkish) with native American English. Our experimental evaluation shows that supervised classification outperforms unsupervised classification by 11.5%. In general, supervised classification performance increases to 80% correct accent discrimination as we increase the phoneme sequence to 11 accent-sensitive phonemes.

Full Paper

Bibliographic reference.  Angkititrakul, Pongtep / Hansen, John H.L. (2003): "Use of trajectory models for automatic accent classification", In EUROSPEECH-2003, 1353-1356.