Japan. Information Science and Technology, The University of Tokyo, Japan. ineering, Information and Systems, University of Tsukuba, Japan. CT enerative adversarial network ion as a discriminator. A ba- ent a real-data distribution by uishes real and generated data. esent the outside of a real-data ception, humans can recognize essed (i.e., a non-existent hu- human-acceptable distribution ne and cannot be modeled by an-acceptable distribution, we enerator training algorithm by ack-boxed discriminator. The training by using a computer . We evaluate our HumanGAN demonstrate that it can repre- Prior distr. Generated data Generator Discriminator Natu- ral Train to fool computer-based discriminator. GAN Training Distr. of training data Generation Distr. of human perception Fig. 1. Comparison of basic GAN and proposed HumanGAN. Basic GAN trains generator by fooling DNN-based discriminator (i.e., computer-based discriminator), and generator finally represents training-data distribution. In comparison, HumanGAN trains gener- ਓؒΛ᱐͢ੜثΛֶश͢Δ͜ͱͰ ਓؒͷײੑΛऔΓࠐΉ ྫɿϞωͱγεϨʔͷֆͷྨ ࣝผثͱͯ͠ਓؒΛ༻͍Δ K. Fujii, Y. Saito, S. Takamichi, Y. Baba, and H. Saruwatari: HumanGAN: Generative adversarial network with human-based discriminator and its evaluation in speech perception modeling. In ICASSP, 2020.
babalablab
honzajavorek
baasie
tonyrice
katarinadahlin
schroneko
sergeychernyshev
aleyda
bethany
jacobian
nikkihalliwell
tammyeverts
jlugia
