Identification of Indian classical languages using Convolutional Recurrent Neural Networks

Language Identification (LID) is a well explored research problem, and it is useful as a front end for several speech processing applications. The task of LID system is to recognize the exact language in a given speech utterance. Several LID systems exist for global languages, trained using several hours of data. But it is a challenging task for Indian languages due to similarities of languages and scarcity of labelled data resources for training. For that reason, this paper aims to solve the LID task of Indian languages using a Convolutional Recurrent Neural Network (CRNN) that uses raw speech as the input. CRNN model is the combination of Convolutional Neural Network (CNN) along with Long-Short-Term-Memory (LSTM). As a result, the CRNN model is assumed to capture the information of both sequential and spatial features from the visual representation of sound. The paper focus on the implementation of a speaker-invariant LID system that classifies all the Indian classical languages namely Kannada, Malayalam, Odia, Sanskrit, Tamil and Telugu. Apart from the identification of six Indian languages, performance comparison of CNN models and CRNN models is also demonstrated in this paper. The LID system also compares with some existing LID techniques and discovers that CRNN model is very effective and significantly outperforms some of the existing LID models.