Fast music information retrieval using PAT tree based dynamic time warping

Recently, music retrieval is getting more and more attentions, especially for the query by humming (QBH), which is straightforward and convenient for usage. However, the difference between humming phrase matching and the precision of singer humming increase the difficulty of retrieval. Apart from retrieval precision, retrieval time is another issue should be taken into consideration. In this paper, we use two-stage searching approach for music retrieval. Since PAT tree (variant of Patricia tree, practical algorithm to retrieve information coded in alphanumeric) shows an excellent performance in matching partial sequence, in the first stage, the simply reduced note interval is adopted as a key phrase of PAT tree for fast indexing the candidate regions from the feature sequences in the database. In addition, the searching of partial sequences suffering from the insertion, deletion and transposition errors works well in the indexing of reduced note interval. The resulting diversity of candidates may be further reduced via a more complicated dynamic time warping (DTW) comparison in the second step. Our approach not only avoids the exhausted computation in the simple DTW approach, but preservers the fault-tolerance capability in the matching process.