Content-Based Recognition Of Musically Similar Music Audio

With the rapid development of multimedia and Internet technology, music is readily accessible in digital form. The amount of music tracks is getting so huge, which leads a great challenge of searching, retrieving and organizing on music content. At present, the most common way of Music Information Retrieval (MIR) is through textual metadata. These human-entered track-level tags reveal the musical and cultural properties, and are extremely difficult to maintain their consistency because anybody can label one track with any description. Meanwhile, Content-Based MIR is engaged in intelligent processing of music, which is thriving nowadays and focuses on automatic music perception and cognition.Besides the intensive research on match between pieces of music, we explore the information about timbre and provide a method to extract the timbre sequence in the music segments, which is full of clues about how the music sounds and proposed to represent the track of music well. So we can solve the matching problem by calculating the similarity between the couple of timbre sequences.First we bring metric distribution of timbre based on the differences and relationships among the feature vectors in varied timbre, which expresses the signal’s characteristic on timbre. Consequently a model for extracting timbre sequence based on GMM and HMM is constructed.And then we bring similarity measurement on timbre sequences mixed sound’s nature and similarity of time series. But the lengths of pieces of music are always not the same, so a newer matching algorithm is introduced using slide-window technique. At this point, the system of matching between pieces of Peking Opera music is built up, in which we can employ various similarity measurement of time series as candidate. A group of evaluations is conducted to indicate it can obtain the result effectively with better performance.