基于非负矩阵分解的钢琴多音符估计  

Multiple Pitch Estimation for Piano Music Based on Non- negative Matrix Factorization

作　　者：万玉龙[1] 王宪亮[1] 周若华[1] 颜永红[1] 

机构地区：[1]中国科学院声学研究所语言声学与内容理解重点实验室,北京100190

出　　处：《网络新媒体技术》2014年第5期23-27,共5页Network New Media Technology

基　　金：国家自然科学基金(批准号:11161140319;91120001;61271426);中国科学院战略性先导科技专项(面向感知中国的新一代信息技术研究;编号:XDA06030100;XDA06030500);国家863计划(资助号:2012AA012503);中科院重点部署项目(编号:KGZD-EW-103-2)经费资助

摘　　要：提出了一种基于能量谱包络非负矩阵分解的钢琴多音符估计算法。首先对钢琴88个单音片段进行RTFI时频分析,求得对应平均能量谱,经过时序平均、归一化求得平均能量谱包络,拼接成钢琴的单音能量谱包络基矩阵。之后对测试的多音片段,采用同样处理方法求得多音平均能量谱包络,通过非负矩阵分解求得各音符的权重系数,最后通过阈值限定求得多音符估计结果。性能评估实验基于MAPS数据集的UCHO集和RAND集展开,与MIREX中最好的钢琴音乐自动记谱系统相比,本文提出的钢琴多音符估计算法性能有很大幅度的提升。A multiple pitch estimation （MPE） algorithm for piano music was proposed here based on non -negative matrix factorization （NMF） of energy spectrum envelope. Firstly, the average energy spectrums （AES） of 88 piano notes fragments are calculated using Resonator Time - Frequency Image （RTFI）. Then the average energy envelopes （ AEE ） are obtained by AES normalization across time. After that, the AEEs are combined to form the average energy envelope basis （AEEB） , which is then used for the NMF of poly- phonic average energy envelope （PAEE）. The weight coefficients of piano notes are calculated from the NMF. Finally, the estimation results are obtained by threshold limitation. Performance evaluation experiments were carried out on UCHO and RAND subsets of MAPS database. Compared with the MPE algorithm used in the best AMT system in MIREX, our proposed one outperforms with better performance.

关 键 词：自动音乐记谱 多基频估计 钢琴音乐 非负矩阵分解 

分 类 号：TN912.3[电子电信—通信与信息系统]