Method of discriminating known and unknown environmental sounds using recurrent neural network

This paper describes our method to classify nonspeech environmental sounds for robots working. In the real world, two main restrictions pertain in learning. First, robots have to learn using only a small amount of sounds in a limited time and space because of restrictions. Second, it has to detect unknown sounds to avoid false classification since it is virtually impossible to collect samples of all environmental sounds. Most of the previous methods require a huge number of samples of all target sounds, including noises, for training stochastic models such as the Gaussian mixture model. In contrast, we use a neurodynamical model to build a prediction and classification system. The neuro-dynamical system can be trained with a small amount of sounds and generalize others by inferring the sound generation dynamics. After training, a self-organized space is structured for the sound generation dynamics. The proposed system classify on the basis of the self-organized space. The prediction results of sounds are used for determining unknown sounds in our system. In this paper, we show the results of preliminary experiments on the proposed model's classification of known and unknown sound classes.