Perceptual features based rapid and robust language identification system for various indian classical languages

Abstract

In this paper, we have investigated the development of the robust recognition system to identify the language of the spoken utterance in several classical Indian languages. The ability of the machines to identify the language in which communication takes place globally is a paramount in the current scenario. In this work on VQ clustering based language identification, feature vectors extracted from training speeches are converted into set of language specific clustering models. During testing, features extracted from test speeches are applied to the language specific clustering models and hypothesized language is identified based on minimum of averages corresponding to the model using minimum distance classifier. Clustering technique is evaluated with variations in cluster size and length of the test utterances. Better performance is observed for the cluster size of 64 and 900 test utterances of 2–3 s duration. Noteworthy point to be mentioned is that though the clustering technique is an old technique, it provides 97% as average recognition accuracy, 3.13% as average false acceptance rate and 3.13% as false rejection rate for the testing done on the language specific clustering model with 64 as cluster size. To improve the accuracy, formants are also used as a feature and it is observed that formant frequencies do provide complementary evidence in ensuring better accuracy for some of the languages. Performance of the system is assessed by calculating the identification rate as the one corresponding to the correct identification with respect to either MFPLPC or Formants and the overall average accuracy obtained is 99.4% for clustering model with 32 as cluster size. Experiments are conducted on the database containing speech utterances in seven classical and phonetically rich speaker specific Indian languages such as Bengali, Hindi, Kannada, Malayalam, Marathi, Tamil and Telugu.