Enhanced spectral features for spoken language identification

To date, systems for the identification of spoken languages have normally used spectral features such as the MFCC. It has also been shown that prosody features such as pitch and intensity have an important role to increase the accuracy of LID. In this paper, we used three novel features based on spectrum, in combination with MFCC and prosody features to improve language identification accuracy. These features are spectral centroid, Renyi entropy and Shannon entropy. The basic system used is GMM-LM with back-end classifier. Also we used a new method to convert the output of the language model scores to the new vectors to increase the LID performance. Using three new features and LM score conversion, an improved accuracy of about 8%, in comparison to the baseline system, on five of the languages of OGI-TS multilingual telephone speech corpus, was obtained.