On a Binaural Model with Front-back Discriminator using Artificial Neural Network trained by multiple HRTF catalogs

Various binaural models have been proposed for the application of hearing assistance system as well as humanoid robot, and a frequency domain binaural model(FDBM) is the one. Like other binaural models, the original FDBM can separate and segregate a signal from the specific direction based on interaural information, but it works only in the frontal semisphere due to front-back confusion. In order to reduce this confusion, a front-back discriminator was proposed for FDBM using artificial neural network (ANN) trained by a head related transfer function (HRTF) catalog. This discriminator has strong dependency on the trained HRTF catalog thus it is not robust against various fluctuation including individual differences and reverberation. This paper discusses an extention of the discriminator using multiple HRTF catalogs for ANN training. The simulation results for the new discriminator show the possibility to reduce the front-back confusion under various conditions including ones obtained in a reverberant room.