Automatic Classification of Indian Languages into Tonal and Non-tonal Categories Using Cascade Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) Recurrent Neural Networks

摘要

This work aims to develop an automatic tonal and non-tonal language classification system of Indian languages using cascade Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) Recurrent neural networks (RNNs). Motivated by their success in modelling sequences, this study proposes LSTM-RNNs cascaded with CNN in context of tonal and non-tonal language classification. Here RNNs show its effectiveness to exploit temporal dependencies in acoustic data. This paper also proposes the use of pitch chroma spectrogram coefficients to address this classification task. The proposed feature is then combined with log-mel spectrogram coefficients to enhance the system performance. The system has been tested for NIT Silchar language database (NITS-LD) which is developed for 9 Indian languages using All India radio broadcast news. And it reports accuracies of 82.30% for 10s and 81.16% for 3s data. Performance of the proposed system is also analyzed on Oregon Graduate Institute Multi-Language Telephone-based Speech (OGI-MLTS) database. It shows accuracies of 77.2% and 74.95% for 10s and 3s data respectively.