Speaker recognition with neural networks.
Masayuki Umino, Shingo Nishimura, Hideki Asoh, Tsunehiko Koike · The Journal of The Institute of Image Information and Television Engineers · 1990
This paper presents the results of speaker recognition (identification and verification) experiments using back-propagation type neural networks.Voice samples of 5 male subjects over a long period of time (1-21 week) are pre-processed and used as input data.First, identification/verification accuracy was examined in several different pre-processing ways, resulting in a 100% correct identification rate and a 94.1% verification rate.Next, the networks performance was investigated by varying the number of hidden units.Some text-independent speaker identification experiments were also carried out.By learning several pre-processed words to decrease the time varying information, networks achieved a correct identification rate of 91.3%.1.