太赫兹科学与电子信息学报, 2019, 17 (3): 482, 网络出版: 2019-07-25
声乐主旋律的自动提取
Automatic extraction of vocal music theme
主旋律 音符切分 维特比算法 基频判别模型 theme note segmentation Viterbi algorithm fundamental frequency discrimination model
摘要
提出一种基于多候选基频提取和歌声基频判别的声乐主旋律提取算法。该算法可以有效降低旋律定位虚警率,提高整体准确率。利用度量距离 (DIS)算法对音乐进行音符切分,并用方差法实现浊音段检测;采用幅度压缩基音估计滤波器 (PEFAC)多基频提取技术,通过计算音高显著度提取每个浊音帧的多个候选基频。最后用维特比算法跟踪浊音段主导基频轨迹,并用基频判别模型进行歌声主旋律判别。在 MIR-1K数据集上进行的实验表明,在信干比为 5 dB和0 dB的情况下,本文算法提取的声乐主旋律整体准确率分别达到了 86.22%和77.4%,相比于其他算法至少提高了3.79%和2.01%。
Abstract
This paper presents a vocal themes extraction algorithm based on multi-candidate fundamental frequency extraction and singing voice fundamental frequency discrimination. The algorithm can effectively reduce the voicing false alarm rate and improve the overall accuracy. First, using the Distance(DIS) metric distance algorithm to achieve note segmentation, and using the variance method to detect voiced segments. Then Pitch Estimation Filter with Amplitude Compression(PEFAC) multi-fundamental frequency extraction technology is utilized to extract multiple candidate fundamental frequencies of each voiced frame by calculating the pitch saliency. Finally, the dominant fundamental frequency trajectory of the voiced segment is tracked by the Viterbi algorithm, and the main melody of the singing voice is determined by the fundamental frequency discrimination model. Experiments conducted on the MIR-1K dataset show that the overall accuracies of the vocal themes extracted by the proposed algorithm reach 86.22% and 77.4%, respectively, at the signal to interference ratio of 5 dB and 0 dB, which are increased by at least 3.79% and 2.01% respectively compared to other algorithms.
陆雄, 夏秀渝, 蔡良, 孙文慧. 声乐主旋律的自动提取[J]. 太赫兹科学与电子信息学报, 2019, 17(3): 482. LU Xiong, XIA Xiuyu, CAI Liang, SUN Wenhui. Automatic extraction of vocal music theme[J]. Journal of terahertz science and electronic information technology, 2019, 17(3): 482.