GMM-Based Identification of Indonesian Speech

This paper reports the performance of identification of Indonesian speech within a ten-language corpus: English, German, Hungarian, Indonesian, Italian, Korean, Mandarin, Polish, Portuguese, and Swedish. The tasks are performed by implementing Gaussian Mixture Model (GMM) on MelFrequency Cepstral Coefficients (MFCCs). Two types of experiments that have been undertaken: pair-wise and tenlanguage experiments. In the pair-wise experiments, the performance of the model in identifying Indonesian within every language pair is evaluated. The experiments show that Indonesian is best distinguished from English, Korean, and Portuguese with 90.5% of accuracy. In ten-language experiments, the highest accuracy of identifying Indonesian is 85.71%.