Research And Implementation On Music Transcription Technology

With the explosive growth of Internet music and the fast development of themultimedia and signal processing technologies, digital music service has already becomeone of the mainstream applications in Internet. Content-based music applications havebrought great changes in musical human-computer interaction, and users can get variousmusic services directly by inputting audio. Music transcription technology is one of thefundamental technologies in the digital music service, which is the linkage of the musicapplication and real music.This thesis focuses on pitch analysis task in music transcription, consisting of audiomelody extraction and multi-F0estimation.(1) As for audio melody extraction task, a newalgorithm is proposed based on spectrum peak and sub-harmonic summation to extractmelody. Firstly, spectrum peak is used to select the F0candidates, then, the true F0isdetected by sub-harmonic summation based salience function. In the comparativeexperiment, proposed method outperforms Cepstrum method, HOD-based algorithm andYIN algorithm in accuracy and anti-interference ability, and is second only to YINalgorithm in speed.(2) As for multi-F0estimation task, this thesis compares and evaluatestwo multi-F0estimation methods: iterative method and joint method. Evaluationexperiment result illustrates: joint method is superior in overall performance to iterativemethod, with better accuracy and adaptability; note number estimation error is a majorcause of multi-F0estimation error, and is an important breakthrough for improvement.The proposed algorithm for audio melody extraction is accurate, fast andanti-interference, and the evaluation result of current multi-F0methods provides theimprovement direction. The research results of this thesis solve pitch analysis problem inmusic transcription effectively, and can be employed in all kinds of music applications.