We study some of the unexplored issues in the parallel phone recognition (PPR) system for automatic language identification (LID). We consider three types of scores for LID, namely, the acoustic score, language model score, and the joint acoustic-language score. Using each of these scores we formulate three types of classifiers for performing LID: maximum likelihood classifier (MLC), Gaussian classifier (GC) and K-nearest-neighbor classifier (KNNC) and compare their performances. We examine the problem of bias in the PPR scores which affects LID performance and interpret the bias and bias-removal methods which improve the classification accuracy of MLC. Among all the different combinations of scoring methods and classifiers, it is found that MLC with bias-removal performs best for either acoustic or language model score alone; this is closely followed by GC with the joint acoustic-language score.
Cite as: Ramasubramanian, V., Sai Jayram, A.K.V., Sreenivas, T.V. (2003) Language identification using parallel phone recognition. Proc. Workshop on Spoken Language Processing, 109-116
@inproceedings{ramasubramanian03_wslp,
author={V. Ramasubramanian and A. K. V. {Sai Jayram} and T. V. Sreenivas},
title={{Language identification using parallel phone recognition}},
year=2003,
booktitle={Proc. Workshop on Spoken Language Processing},
pages={109--116}
}