Two Stage Neural Network model for Recognition of Indian Languages from Speech

India is a multilingual country. Officially about 20 languages are recognized by the government and there are about 500 languages spoken at different parts of the country. For developing the speech systems in Indian context, it is necessary to capture the language specific knowledge automatically from speech. Further it may be exploited for different speech tasks such as language identification, speaker identification, speech recognition and so on. In this paper we focus on identification of 15 Indian languages from speech using spectral features. In this work, spectral features are represented by Mel frequency cepstral coefficients. The 15 Indian languages considered for this study are: Assamese (As), Bengali (Ben), Gujarati (Guj), Hindi (Hi), Kannada (Ka), Kashmiri (Kas), Malayalam (Mal), Marathi (Mar), Nepali (Nep), Oriya (Or), Punjabi (Pun), Rajasthani (Raj), Tamil (Ta), Telugu (Te) and Urdu (Ur). Here neural network models are explored for capturing the language specific information from the spectral features. Its known that the vocal tract characteristics are different, for producing the sound units in different languages [1]. Directly attempting to develop simple language models using neural networks may not give the optimal performance, as most of the Indian languages are originated from the single ancient root language Sanskrit, and hence they are confusing. The degree of confusion is very high in the languages of neighboring geographical regions.