Autonomous voice acquisition of a talking robot based on topological structure learning by applying dual-SOM

This paper presents the method of autonomous voice acquisition for a talking robot by applying a dual-SOM. We have so far developed the talking robot, which has mechanical organs as a human. By applying an auditory feedback control, the robot autonomously learns the vocalization skill. For the autonomous learning method, a Self Organizing Neural Network (SONN) by combining a Self-Organizing Map (SOM) with a Neural Network (NN) was employed. The SONN had 2-dimensional mapping space, which was used to locate phonetic features of voices generated by the robot. By choosing cells on the map, voice articulations were autonomously recreated. However, due to the spatial restriction of the map, the voice transition from one vocal sound to another was not always recreated properly. To solve the problems, a dual-SOM, the combination of a phonetic-SOM and a motor-SOM, both of which have 3 dimensional mapping spaces, is introduced. The structure of the dual-SOM is firstly described, and then acquired vocal sounds are evaluated, together with the analysis of the behavior of the SOM.