Automatic language identification with recurrent neural networks

Automatic language identification (LID), an important domain in speech processing, means the capability of a machine to determine a natural language from a spoken utterance. We present a novel approach to LID, which involves recurrent neural networks (RNN) as the main mechanism. We propose that, because of acoustical context issues, RNNs are particularly suitable for the LID task. Our approach also introduces perceptually guided training (PGT), a novel training method, that exploits the concept of perceptually significant regions (PSR) postulated by our approach. We present our overall approach and describe LIREN/PGT, the have we have developed, implementing our approach. We also discuss our LID experiments with English, German, and Mandarin. In the paper, we concentrate on the architecture and training of RNNs for automatic language identification.