Speaker identification using modular recurrent neural networks

This paper demonstrates a speaker identification system based on recurrent neural networks trained with the Real-time Recurrent Learning algorithm (RTRL). A series of speaker identification experiments based on isolated digits has been conducted. The database contains four utterances of ten digits spoken by ten speakers over a period of nine months. The results suggest that recurrent networks can encode static and dynamic features of speech signals. They also show that the proposed system outperforms the traditional speaker identification systems in which Backpropagation networks are used. However, this paper demonstrates experimentally that the outputs of the RTRL networks are highly dependent on the initial portion of the input sequences. Removing the first few vectors from the input sequences will lead to a substantial reduction in identification accuracy.