Language identification system for South African languages

In our work to design a speech language translator between South African languages we found the project inflexible and much bigger than first anticipated. The first part of the project was to design a system that will perform language identification. This paper investigates a language identification (LID) task using three South African languages. The problem is investigated solely from the signal processing perspective using linear predictive coding (LPC)-based and parameterized discrete Fourier transform (DFT)-based feature-sets. Six minute speech data was collected from a talker in all three languages. The LID system uses five seconds samples of speech in the three languages to perform identification. The results show that a DFT-based parameterized feature-set significantly lowered the error rate. The lowest error rate is obtained at a spectral compression lower than the mel-scale.