| The exploding corpus of digital audio files on Internet has led to unprecedented convenience in sharing and appreciating musical works. Following this trend, almost all famous Internet corporations have been engaged in music service provision. In this way, those research topics on music information analysis have become more and more popular in recent years.Almost all early methods for music information analysis and retrieval heavily depend on manual annotations on all music data. However, with rapidly growing size of music database, to handle all songs in this way has become more and more time-consuming and expensive. To conquer the weakness of manual work, automatic music data analysis methods spring up in these years. Such topics include Cover Song Identification, Audio Matching, Music Structure Analysis, Humming system etc. Usually, such methods do not require human annotations, taking the advantage of signal-level features directly extracted from raw musical works.Among all those research topics, Similar Segment Detection for audio pieces is a very important sub-topic as well as fundamental technology to support other researches. Similar segments among musical works are ubiquitous, for example, chorus pieces in a song usually share same melody contour. Generally two audio pieces are announced to be similar because they share similar melodies, rhythms or timbre etc. Such similarities are not difficult for human to recognize, while they remain unresolved problems for computer processing. Due to the fact that those features directly extracted from songs contain huge noise, previous similar segment detection strategies are far from perfect and mostly suffer from following defects: firstly, the detection accuracy is low; secondly, the positions of similar segment pairs cannot be detected accurately; thirdly, short similar pieces cannot be handled by current methods.To handle above mentioned problems in previous methods, this paper proposes a new similar segment detection strategy in music works. The newly designed method integrates a variety of existing technique as well as self-designed algorithms. This method is then separately applied to three sub-tasks: cover song identification, music summary extraction and music borrowing. Among those tasks, music borrowing is a new research topic, which has no specific research works on it yet. This topic is firstly explored in this paper due to the capability of the proposed algorithm to detect short similar segments in musical works. Experiments on the three systems show the effectiveness and improvements of the proposed methods. |
