7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Stochastic Trajectory Model Analysis for Accent Classification

Pongtep Angkititrakul, John H. L. Hansen

University of Colorado at Boulder, USA

It is believed that knowledge gained from reliable accent classification could be employed to improve the performance of speech recognition and speaker recognition algorithms. In this paper, we investigate the use of articulatory movement in the spectral domain to classify accented speech. The trajectories are modelled by a mixture of probability density functions of a random sequence of states. The approach is based on a Stochastic Trajectory Model (STM) which has been considered for speech recognition[1] and speech synthesis[2]. The CU-Accent database is collected over a telephone channel using speakers with foreign language accents. Experiments are performed at a context-independent phoneme-class level. Accent classification evaluations using English produced by native speakers of Mandarin Chinese, Thai, Turkish, and native American English showed classifi- cation rate in the range of 64.2-67.4% for STM versus 62.4-66.6% for GMM, using single phonemes. One of the key results is the formulation of an accent sensitive phoneme tree across the four English accents.

Full Paper

Bibliographic reference.  Angkititrakul, Pongtep / Hansen, John H. L. (2002): "Stochastic trajectory model analysis for accent classification", In ICSLP-2002, 493-496.