基于谐波结构信息的自动音乐标注方法

Automatic Music Transcription Based on Harmonic Structure Information

根据同类乐器音色相似的特点,提出了一种基于谐波结构信息的自动音乐标注方法·该方法事先提取一类乐器中某一件乐器的谐波结构信息·根据输入信号选择合适的不谐和系数、频率误差系数,结合谐波结构信息构造同类未知乐器的响度谱,然后采用截断完全最小二乘法实现对同类未知乐器音乐的标注·对钢琴音乐的实验证明,该方法不仅能对未知乐器的音乐进行标注,取得较好的性能,还可辨别音符的响度强弱·

A novel method for automatic music transcription based on harmonic structure information is developed in this paper. The instruments of the same class have similar timber and harmonic structure. Therefore, harmonic structure information of an instrument is acquired beforehand and is used to construct the loudness spectrum for the unknown instrument of the same class in music transcription. The inharmonicity factor and the error in note fundamental frequency are considered during the spectrum construction. The loudness spectrum of input audio can be considered approximately as the linear sum of spectra corresponding to the mixed notes and the truncated total least squares is used to solve the sum equation. Experimental results of piano music transcription show that the method not only can transcribe music of an unknown instrument with high performance, but also can distinguish the loudness of mixed notes.