Retrieving Vaguely Remembered Lyrics Using N-Gram Edit Distance

Current text based music information retrieval systems are based on full-text retrieval engines or matching the exact keywords. If a user vaguely remembers lyrics, those systems are incapable of searching for lyrics. The major type of vaguely remembered is spelling variants. In this paper, we propose using kana for the retrieval of lyrics where queries may contain spelling variants. First, we construct a standard inverted index over the kana converted lyrics. Next, we filter the search result using n-gram Levenshtein distance. We demonstrate the effectiveness of the system through an experiment using queries containing one, two or three spelling variants. From the experiment, all accuracy rates were higher than 90% when the query contains one, two or three spelling variants.