一种基于隐马尔科夫模型的波形文件主旋律基频提取算法

摘要

Query-by-Humming (QBH) systems generally use the pitch contour of melody as music features. The present QBH systems are all based on MIDI files. However, the quantity of existed MIDI files is very small. The QBH systems based on waveform ifles will be the future trend. This paper introduced a melody pitch extraction algorithm for monaural recordings. The algorithm proposed a combination of Hidden Markov Model(HMM) and Harmonic/Percussive Sound Separation (HPSS). Experimental results show our algorithm performs excellent accuracy.%哼唱检索中通常以旋律的基频作为音乐特征进行检索，目前研究的哼唱检索系统都是基于MIDI音乐文件。但是，目前存在的MIDI音乐文件的数量非常少，基于波形文件的哼唱检索系统才是未来的发展趋势。本文针对单声道波形文件，研究了一种提取歌曲主旋律基频曲线的算法。该算法将隐马尔科夫模型和“谐波乐器/打击乐器声音分离”模型进行结合。实验表明该算法对主旋律基频提取具有很高的准确率。