A Comparative Study of Feature Modelling Methods for Telugu Language Identification

Telugu is one of the most widely spoken language, with 82.7 million native speakers. Telugu is used predominantly in the Indian states of Telangana and Andhra Pradesh. This paper explains the design and implementation of a model that identifies whether the language spoken by the user is Telugu or from a multilingual dataset containing audio clips from Telugu, Malayalam, Gujarati, English, and Marathi of native Indian speakers. Five features, MFCC, Chroma CQT, Chroma STFT, Chroma CENS, Tonnetz-space, and combinations of these, are used for training the models. Five powerful deep learning architectures CNN, LSTM, CNN-LSTM, BiLSTM, and GRU are used in training over the features and different feature combinations. The performance of the models has been compared with state-of-the-art models for language identification. The proposed models yielded maximum accuracy of 94% for CNN with Chroma CQT+MFCC feature combination.