An estimation of speaker sampling in Voice Across Japan database

This paper discusses a speaker-sampling problem in data collection. In collecting a huge speaker-corpus, the sampling of the speakers becomes a very important problem. In the Voice Across Japan (VAJ) project, we have gathered 8,773 speakers´ data through a telephone line in proportion to population as possible as we can, and also have collected the speakers´ information such as gender, age, and growing area. We discuss the problem where the parameters of the speakers such as, the “age” or “growing place” influence the speech recognition. If such parameters are not so important, both data collection and product realization will be easier. In our experiments, it was clearly found that the “speakers´ age” is the strongest factor and that the “growing area”, also, have an influence in the range of 2-4%