Key Estimation Using a Hidden Markov Model
Key Estimation Using a Hidden Markov Model
复制标题
使用隐马尔可夫模型的关键估计
DOI：
--
复制
发表时间：
2006
期刊：
影响因子：
--
通讯作者：
M. Sandler
中科院分区：
文献类型：
--
作者：
K. Noland;M. Sandler
提出了一种新的技术来估计音乐摘录中的主导键。键空间由24状态隐马尔可夫模型（HMM）建模，其中每个状态代表24个主要和次要键中的一个，每个观察代表一个和弦过渡或一对连续和弦。使用和弦过渡作为观察模型的连续和弦之间的时间依赖性比将观察单个和弦。关键的过渡和和弦发射概率初始化使用的感知测试的结果，以反映人类的期望的谐波关系。然后，在解码每首歌曲的模型以给出每个时间帧每个键的可能性之前，使用手工注释的和弦符号在每首歌曲的基础上训练HMM参数。作为一个分割技术的算法的例子给出，其能力来估计一首歌的整体关键的评价使用的数据集的110披头士歌曲，其中91%被正确分类。扩展到包括从音频数据而不是和弦符号的操作，这将使应用程序的一般音乐检索的目的。
A novel technique to estimate the predominant key in a musical excerpt is proposed. The key space is modelled by a 24-state Hidden Markov Model (HMM), where each state represents one of the 24 major and minor keys, and each observation represents a chord transition, or pair of consecutive chords. The use of chord transitions as the observations models a greater temporal dependency between consecutive chords than would observations of single chords. The key transition and chord emission probabilities are initialised using the results of perceptual tests in order to reflect the human expectation of harmonic relationships. HMM parameters are then trained on a per-song basis using handannotated chord symbols, before the model for each song is decoded to give the likelihood of each key at each time frame. Examples of the algorithm as a segmentation technique are given, and its capability to estimate the overall key of a song is evaluated using a data set of 110 Beatles songs, of which 91% were correctly classified. An extension to include operation from audio data instead of chord symbols is planned, which will enable application to general music retrieval purposes.