Text‐independent talker identification using recurrent neural networks

Abstract

Recurrent neural networks (RNNs), i.e., artificial neural networks with internal feedback, have the potential to perform pattern classification using both static and dynamic spectral information. Experiments were conducted to determine if these networks can be trained to identify talkers in a text‐independent mode using short samples of speech for training the network. The database for experimentation consisted of the first 200 ms of 60 CVC syllables for each of 30 speakers. Ten of these speakers were adult males, ten were adult females, and ten were children. Thirty syllable segments (6 s of speech) were used for training and 30 different syllable segments (6 s) were used for testing. To speed up the training times for the RNN, four independent RNNs were used. One RNN was used to differentiate the speaker type (man/woman/child) and each of the other networks was trained to differentiate among the ten speakers in each group. For each RNN, ten input nodes (corresponding to ten modified cepstral coefficients) and 30 hidden nodes were used. With test words for evaluation, the RNNs were able to identify the speaker group for all speakers (i.e., 100% accuracy). The RNNs also identified the individual speakers with 100% accuracy. These results indicate that recurrent neural networks can be successfully trained in a text‐independent talker identification task, even with a relatively small amount of data.