We, people, recognize our surroundings with five senses. Out of these five senses, it is generally considered that the main one is the visual sense. However, it still has some week points, such as visible light is not effective when there is a screen, or even the recognition itself can not be made, if there is no light source. Compared to the visual sense, the auditory sense can avoid the above problems naturally. Although it is behind the visual sense in terms of the resolution, the auditory sense also play an important role in recognizing surroundings obviously [1, 2].We have focused on this information processing mechanism of the auditory sense, and aimed at the construction of a surroundings-recognition system for artificial objects like automotive robots and vehicles, etc. It has been constantly studied to apply neural networks to information-processing systems related to sound these days. The networks that have been used in those studies are mostly based on the McCulloch-Pitts type neuron model [3]. Regarding the neuron models, they have many advantages of having well-established learning rules for multi-layered networks [4], and so on. On the other hand, there is also a problem about that kind of models, such as that the data of the inputs and outputs shows static values, and the neurons themselves do not have dynamics. Due to this problem, even if the temporal variations of signals carry crucial information, these signals have to go through the (so-called unnatural) windowing process along time axis, and they must be dealt with as the multi-dimensional vectors.In contrast, what we have already proposed is the temporal information processing system using pulsed neuron models [5, 6, 7]. Pulsed neuron models (* described as PN models in the rest of this paper for convenience) are the neuron models in which pulses are regarded as the input and output signals. In the neural network using PN models, temporal input signals are converted into the pulse trains, in which the information of these signals' intensity is kept as the pulse frequency. Therefore, there is no necessity of the windowing process for the PN models. Furthermore, the PN models can be constructed of simple analog or digital circuits, and each model can work independently and asynchronously.In our study until now, these PN models have been applied to the sound localization model [5, 6] In this models, when the pulse trains are inputted to the PN models that are assembled variously, it is possible to detect the specific sound direction. In addition, we have proposed the supervised learning rules for the PN models [7] that have been useful for the above model. As I have said, PN models and the network using them are effective in processing the temporal information. However, when it comes to giving input signals directly to the PN models with the supervised learning rules, it is not reasonable because each pulse train of input signals varies its pattern frequently and also the volume of the data is enormous. Therefore, before the supervised learning rules are carried out, the information of the input signals needs to be compressed in some ways.Accordingly, in this paper, we propose the unsupervised learning rules and the method of the vector quantization for the PN models to compress the temporal information per every instantaneous time. In the current neural networks, the unsupervised learning rules are widely employed for the vector quantization, dimensionality reduction, self-organization, etc. For the prospective application of the PN models, it is significant to establish the unsupervised learning rules for the models. In terms of the unsupervised learning rules, we examine the application of Kohonen's Competitive Learning Algorithm [8].