| Music, the most ancient but prevalent form of art, is now existing as"net music"in a massive, inorganized, scattered state, along with the boom of internet. As music is now still closely and inseparably associated with our everyday life, Music Information Retrieval(MIR), a technique focuses on effectively searching and browsing on net music sources and databases, is becoming research hotpot. On the other hand, in the field of music teaching and composition, increasing need for computational and intelligent application of musicology to improve outdated teaching and composition methods, makes study of musicology based music intelligent computing technique imperative.One core problem of the above two techniques is, how music audio can be intelligently analysed so that elements(basic and formal elements which compose music structure and expression methodology respectively) of music content will be learned easily. Due to research status quo and actual demand of music audio processing, this dissertation regards music elements analysis as key technology in this domain; furthermore, this dissertation presents a novel music processing framework—Music Computing framework, aiming at intelligent and automatic recognition of various music elements, and effective use of these information in music analysis, retrieval, and transcription. To achieve core goal of Music Computing framework, this paper addresses a close combination of musicology and computation in methodology, and tries to solve the problem of music signal analysis and recognition with theories from musicology, cognitive psychology, music cognition, neurology, artificial intelligence, and signal processing. This dissertation investigates into several critical problems in this research field as following:1) This dissertation addresses a 4-level hierarchy of Music Computing framework with theory, method, task, and application layer for the first time. In theory layer, this paper proposes a method of music connotation analysis: dividing connotation of music into sound, rational cognitive, and irrational cognitive dimension, and explains the theoretic fundamental of Music Computing framework with this connotation structrue. In method layer, this paper divides methodological domain into 8 subfields with the combination of music connotation structure and hierachical classification of music element(high-, mid-, low-level), and demonstrates the key research fields of this dissertation. In task layer, this paper classifies goals of research tasks with the music element hierachical classification in method layer, and presents relations between tasks. In application layer, 2 anticipative application direction of Music Computing framework and related key techniques are presented.2) For the study of music signal processing method, this dissertation proposes a Pitch Tuning(PT) technique based on investigating into classical techniques such as Constant Q Transform(CQT) and onset detection, and PT can deal with the situation that tone sounded by instrument deviates from its standard frequency caused by change of instrument's physical conditions. Besides, this paper proposes a Time Accumulated Neural Networks(TANN) with time delayed and accumulated unit, which offers an effective solution to temporal pattern recognition problem; thus, TANN is fairly capable of processing musical feature sequence extracted to recognize certain music element. And a networks training method called Entropy Error Function(EFF) is also demonstrated to enhance the networks'ability of convergence, generalization and efficiency.3) For the study of analysis of music tonal attributes, this dissertation presented a PCDM feature which describes the characteristics of tonality properly, based on music distributional cognition hypothesis. The PCDM, not only describing spectral character of pitch class, but also characterizing spectrum of semitones within octave, is a novel computational cognitive feature which integrates both distributional and structural cognition view. Then PCDM is applied to key and chord recognition in real world music audio, which also involve many music signal processing techniques.4) This dissertation also proposes a novel music genre classification method with a fusion of long-term and short-term feature. A mixed feature set of timbre feature such as MFCC and rhythm feature beat histogram are extracted to classify genre based on Gaussian Mixture Model(GMM). Furthermore, a genre vector based radar chart denotation is presented to offer visualization of classification results—radar chart presentation function well when degree of confusion between genres and music with inter-genre character are studied. Besides, a Support Vector Machine(SVM) based classification method between vocal and melody is studied, with the aid of background knowledge of music structure.5) Finally, based on the above studies and achievements on human perception of music in brain science and cognition, this dissertation proposes an acoustical measurement of AS(AAS) for detecting salient structure of polyphonic music from the point of view of selective attention theory. The AAS is a 3-dimension vector that describes structural and temporal difference between adjacent semitone channels through statistics model. Then this paper proposes a melody part recognition method based on AAS, with a result presentation of melody stream which is closer associated with actual human sense of listening and perception goal. Then Musicological measurement of AS(MAS) is proposed as the extention of AAS with background knowledge of musicology and famous cognitive models by taking musical context relationship and harmonic structure into consideration, and it is also a proof of our theory of music connotation. At last, this dissertation presents a automatic music transcription method with the full version of AS feature,using pour previous work on techniques(CQT, PT), tool(TANN), results of key finding and chord recognition, and regulations of musical rules, achieving a primary comprehensive content analysis of music audio.
|
PDF Full Text Request