Deep Learning Based Language Identification System From Speech

The paper outlines a language identification system from speech with the help of deep learning. In a multilingual country like India, language identification systems have a wide variety of applications. For this, a custom speech database of seven Indian languages from IIIT-Indic has been taken. Preprocessing is done in order to avoid the noise interference and the disturbances in the audio signal. After subsequent framing and windowing, the spectrograms are extracted and fed to a Convolutional Neural Network (CNN). Utterances in languages: Bengali, Hindi, Marathi, Malayalam, Kannada, Tamil and Telugu are utilized in this study. A mixture of speech utterances from binary combination of these languages is used in this analysis. The performance of the system is measured in the form of accuracy, precision, recall, F1 score and confusion matrix. For the data set analyzed, a maximum accuracy of 99.5% is achieved for Bengali vs Kannada and 99% for Kannada vs Malayalam.