Data collection of Japanese dialects and its influence into speech recognition

Reports the successful completion of a Japanese POLYPHONE project-the Voice Across Japan (VAJ) data collection project. The database has the following characteristics: (1) a large speaker database (8,866 speakers) through a telephone line, (2) gathering of the participants' personal information such as gender, age, place where they grew up, and so on, and (3) data segmented by phone or word boundaries. This paper describes several aspects of Japanese dialects and also reports the results of experiments. How much does dialect influence speech recognition? In our results, dialect influences the speech recognition rate by 2-4%. The results are useful information for building practical speech recognition systems as well as for data collection.