Automatic Music Read Music Notes In The System Detection And Genre Classification

This thesis is mainly concerned with how to realize automatic music transcription. Transcription means that the music listener records scores of music he is listening to; the so-called automatic music transcription indicates that computers intelligently record those scores without human interference, during which the most important information recorded include the notes and instruments utilized.The primary content of our work is relevant to note and instrument recognition:Firstly, after comparing five classical algorithms of pitch detection, the thesis proposes a pitch extraction algorithm based on psychoacoustics. After the five algorithms are carefully analyzed and compared, that is HPS (Harmonic Product Spectrum) algorithm, Cepstrum algorithm, CBHPS (Cepstrum Biased HPS) algorithm, Maximum Likelihood algorithm and at last Autocorrelation algorithm, a new one needs being proposed to resist the noise as to promote the accuracy of pitch detection and to satisfy the requirement of efficiency of large scale music analysis systems. Therefore, one method is put forward, which applies the efficient Autocorrelation methods to critical band simulating human's ear, and which provides both high efficiency and detection accuracy.Secondly, to improve the accuracy of note recognition, we propose our pitch-timing revision algorithm. Because in current music analysis domain, no particular detection algorithms are designed for note, but for general onset detection, which always fail to understand the notes or localize them exactly; thus one algorithm is proposed to detect notes more reliably, based on the mutual revision between the pitch and the timings, and so called "inter-revision" algorithm.Thirdly, combining the bottom-up and up-bottom principles, and in order to assist instrument recognition through the information from music genre, this thesis constructs one efficient Bayesian network using eight decisive features, after careful testing all the frame features and segment features. Our Bayesian network classifier achieves much higher accuracy that popular Neural Network classifiers. Meanwhile in our experiment 480 songs from six genres, named Jazz, Rap, Blues, Rock and Roll, Country Music, Cha Cha, are used, as means generality of our method.