基于序列模型的音乐词曲匹配度智能评估算法

PDF下载HTML

Title:

Music lyrics-melody intelligent evaluation algorithm based on sequence model

作者:

陈壮豪¹, 张茂清¹, 郭为安², 康琦¹, 汪镭¹

1. 同济大学 电子与信息工程学院, 上海 201804;
2. 同济大学 中德工程学院, 上海 201804

Author(s):

CHEN Zhuanghao¹, ZHANG Maoqing¹, GUO Weian², KANG Qi¹, WANG Lei¹

1. College of Electronics and Information Engineering, Tongji University, Shanghai 201804, China;
2. Sino-German College of Applied Sciences, Tongji University, Shanghai 201804, China

关键词:

音乐词曲; 情感; 节奏; 序列模型; 歌词编码器; 旋律解码器; 词曲匹配解码器; 词曲匹配度

Keywords:

music lyrics-melody; emotion; rhythm; sequence model; lyrics encoder; melody encoder; matching decoder; lyrics-melody matching degree; music lyrics-melody matching

分类号:

TP393.04

DOI:

10.11992/tis.202001006

摘要:

情感匹配模型是一种常用于评价词曲匹配程度的方法；然而，单纯地依靠情感匹配模型无法对评价词曲匹配度进行准确的评价。为改善此问题，提出了基于序列模型的词曲匹配度智能评估算法，其综合考虑词曲情感和词曲间节奏关系以给出一个更加准确的词曲评估方法。基于公开词曲同步数据集，构建了音乐情感和节奏正反例模型，并基于此模型将音乐切分成片段；进一步，将歌词和旋律片段分别通过歌词编码器和旋律编码器进行编码，并将编码后具有上下语境的歌词特征和旋律特征输入词曲匹配解码器，解析词曲间特征关系，判断词曲片段匹配程度。仿真结果表明：基于序列模型的词曲匹配度智能评估算法，相对于单纯的情感匹配模型，能够更精确地评价词曲匹配程度，验证了本文提出算法的有效性。

Abstract:

Emotional matching model is a method often used to evaluate the degree of lyrics and melody matching. However, it cannot be accurately evaluated based on the emotion matching model. In order to improve it, this paper proposes an intelligent evaluation algorithm of lyrics-melody matching based on a sequence model, which comprehensively considers the emotion and the rhythm relationship between lyrics and melody to give an evaluation method for more accurate evaluation. Firstly, this paper researches and builds music positive and negative samples considering music emotion and phrase based on the public lyrics-melody paired dataset and divide songs to music pieces. Further, the lyrics and melody fragments are encoded by the lyrics-encoder and the melody-encoder, respectively. And take the encoded lyrics feature and melody feature that are contextualized as the input of the lyrics-melody matching decoder to analyze the characteristic relationship between the lyrics and melody, and then determine the matching degree of the lyrics-melody segment. The experimental results show that the music lyrics-melody matching intelligent evaluation algorithm model based on sequence model can more accurately judge the matching degree of lyrics-melody matching than simple music emotion matching, which verifies the effectiveness of the proposed algorithm.