A pitch detector for application in query by humming (QBH) is implemented in this paper. This algorithm is made up of two parts: note segmentation and pitch detection. In the first part, voiced/silence decision is mad...A pitch detector for application in query by humming (QBH) is implemented in this paper. This algorithm is made up of two parts: note segmentation and pitch detection. In the first part, voiced/silence decision is made on each segment of the input signal by a pattern recognition approach, and further, the preparatory note borders are obtained; then, via analysis of the instantaneous energy contour, the adjacent notes that adhere to each other are separated. In the second part, pitch is estimated for all frames contained in a note's duration by an autocorrelation method and the mean of these pitch values is taken as the average pitch of this note. Moreover, in order to remove the effect of formant structure, a nonlinear preprocessing is adopted in the pitch detection part and the autocorrelation function is properly weighted before peak picking. Finally, hummings of several experimenters with different voice characters are recorded to test this pitch detector, whose efficiency and reliability are proved by the results.展开更多
提出了一种近似旋律匹配(approximate melody matching)的新方法——线性对齐匹配法,并在此基础上实现了一个哼唱检索(query by humming)系统原型.与已有的基于内容的音乐检索(content-based music retrieval)不同,该算法并非基于近似...提出了一种近似旋律匹配(approximate melody matching)的新方法——线性对齐匹配法,并在此基础上实现了一个哼唱检索(query by humming)系统原型.与已有的基于内容的音乐检索(content-based music retrieval)不同,该算法并非基于近似符号串匹配、统计模型或者特征空间,而是根据相近旋律的音高轮廓在几何上的相似性,将音高和节奏特征一并考虑所设计而成的全新算法.通过实验检验该算法的有效性,在含有3864首乐曲的搜索空间中,检索62段人声哼唱,线性对齐匹配法取得了90.3％的前3位命中率,相比传统的近似符号匹配算法高出11％以上.这一实验结果有力地表明了线性对齐匹配法的有效性,及其应用于大型数字音乐检索引擎的可行性.展开更多
文摘A pitch detector for application in query by humming (QBH) is implemented in this paper. This algorithm is made up of two parts: note segmentation and pitch detection. In the first part, voiced/silence decision is made on each segment of the input signal by a pattern recognition approach, and further, the preparatory note borders are obtained; then, via analysis of the instantaneous energy contour, the adjacent notes that adhere to each other are separated. In the second part, pitch is estimated for all frames contained in a note's duration by an autocorrelation method and the mean of these pitch values is taken as the average pitch of this note. Moreover, in order to remove the effect of formant structure, a nonlinear preprocessing is adopted in the pitch detection part and the autocorrelation function is properly weighted before peak picking. Finally, hummings of several experimenters with different voice characters are recorded to test this pitch detector, whose efficiency and reliability are proved by the results.
文摘提出了一种近似旋律匹配(approximate melody matching)的新方法——线性对齐匹配法,并在此基础上实现了一个哼唱检索(query by humming)系统原型.与已有的基于内容的音乐检索(content-based music retrieval)不同,该算法并非基于近似符号串匹配、统计模型或者特征空间,而是根据相近旋律的音高轮廓在几何上的相似性,将音高和节奏特征一并考虑所设计而成的全新算法.通过实验检验该算法的有效性,在含有3864首乐曲的搜索空间中,检索62段人声哼唱,线性对齐匹配法取得了90.3％的前3位命中率,相比传统的近似符号匹配算法高出11％以上.这一实验结果有力地表明了线性对齐匹配法的有效性,及其应用于大型数字音乐检索引擎的可行性.
文摘基音检测是音频分析和基于内容的音乐检索中的关键技术,是基于内容音乐检索中实现哼唱检索的基础。提出一种改进的自相关函数(autocorrelation function,ACF)方法进行基音检测。从对音频信号进行去噪预处理、清浊音判断及后处理等方面对ACF进行改进,使之能够生成规整的音高变化曲线。在音乐检索的实现中,提出一种有限长度的最长公共子序列(Local Longest Common String,LLCS)方法,该方法可有效解决传统方法存在的误检问题。开发实现了一个通过哼唱/歌唱进行歌曲检索的原型系统。对大量的歌曲哼唱的实验表明,提出的改进ACF算法和LLCS算法对于提高检索正确率是正确有效的。
