基于遗传算法的旋律匹配模型研究

Melody matching model based on a genetic algorithm

基于内容音乐检索(MIR)以其简捷、直观的检索方式成为模式识别、信号处理等领域研究的热点之一.在MIR问题中,特征的选择、表示和匹配是核心技术.在研究分析音乐物理及感知特征的基础上,以旋律作为主要特征,通过基音提取和动态阈值分割音符算法,为待检索音乐数据集和输入音乐样本建立了旋律表示模型;应用遗传算法对齐模板,修正哼唱输入个体差异,以提高检索精确度;融合欧式距离和动态时间扭曲(dynamic time warping,DTW)相似度实现度量匹配模板,以加强容错和泛化能力.实验表明,算法的速度及精度可良好地满足哼唱系统的要求,并可扩展应用于相似的系统.

Content-based music information retrieval （MIR） has received increasing attention in the field of signal processing and pattern recognition. Methods for feature selection, representation and matching are three key points in the MIR process. Some physical characteristics of musical features are favored in the perceptions of listeners. As a result, melody is thought to be the most important feature. Pitch tracking and dynamic threshold note segmenta- tion were used to formulate a template for queries about music. A melody contour alignment algorithm based on a genetic algorithm （GA） was proposed, which was used to align and correct the input pitch template. A weighted combination of the Euclidean measure and the dynamic time warping （DTW） algorithm was used to make the search process fuzzy and more effective. The experiment demonstrated that the algorithm~ tacking speed and accu- racy could meet the needs of a query-by-humming system, and could be used in similar systems.