ABSTRACT
Real-time language identification is a problem well-suited to neural network solution. By limiting the amount of speech processing at the frontend this can be achieved to a surprising degree. In our experiments we have developed a recurrent neural network which has been trained and tested using examples from English and French. The recurrent network produces a series of activation patterns (one outcome for each cycle) which can be seen as votes for or against each language. These votes are summed over an interval of speech and the segment is classified.
