Abstract:

In the research of instrument recognition and analysis of mixed instrument music data, the existing methods rely heavily on data labels, while are often based on simple frequency domain or physical characteristics, which are not obviously related to the inherent properties of the instrument, and lack of sensitivity to complex components. This paper proposes a sparse feature extraction method based on sparse decomposition and multiple instrument component dictionaries, which can get sparse features that will be used independently & with high interpretability through in-depth analysis of the sparse coefficient vector. Experimental result shows that these features can express the composition of musical instruments and the changes of musical mood directly, and also shows significant application value in the field of composition analysis of mixed musical instruments and other kinds of time-varying signal analysis.

Key words: feature extraction, sparse decomposition, sparse feature, mixed instrument recognition, music time domain analysis

摘要：

为了解决混合乐器音乐数据的成分识别与解析研究中，现有方法过度依赖数据标签，且往往基于单纯频域或物理特征，与乐器固有性质关联不明显、对复杂成分的敏感度不足的问题，提出了一种基于稀疏分解和多种乐器成分字典的稀疏特征提取方法，通过对稀疏系数向量进行深入分析，得到可以独立使用，具有高可解释性的稀疏音乐特征。实验结果证明，这种特征能够直观地反映乐器成分组成与音乐情绪的变化，在混合乐器成分分析和其他各类时变信号分析领域具有显著的应用价值。

关键词: 特征提取, 稀疏分解, 稀疏特征, 混合乐器识别, 音乐时域分析