摘要
指法是钢琴演奏的关键技术,但是除了初学者的教科书外,大多数乐谱都没有指法注释。目前用于钢琴指法自动生成的隐马尔可夫模型(hidden Markov model,HMM)和长短时记忆网络(long short-term memory,LSTM)模型,仅针对乐谱的音高建立模型,忽略同样影响指法的速度信息,存在对乐谱综合特征提取能力不足、生成的指法正确率低等问题。针对这些问题,设计一种可以同时利用乐谱的音高信息与速度信息的特征提取方法,并引入Word2Vec-CBOW(continuous bag-of-words)模型得到融合特征向量,根据人体左右手镜像对称的特点对原始数据进行左右手序列的数据增强与联合训练,最后结合双向长短时记忆网络-条件随机场(bidirectional LSTM conditional random field,BiLSTM-CRF)模型实现指法的生成。实验结果显示,本文提出的算法相比常用的统计学习方法和深度学习方法均有明显提高,验证了其合理性和有效性。
Fingering is a key technique in piano playing.However,most musical scores have no finger notation except in beginners’textbooks.The HMM and LSTM models used for automatic piano fingering only model pitch information and ignore speed information,which will influence the fingering.This condition results in insufficient extraction of comprehensive features and a low accuracy rate for generated fingerings.A feature extraction method was first designed using the pitch and speed information of the musical score simultaneously to address these problems.The Word2Vec-CBOW model was then introduced to produce a fused feature vector.Further,data enhancement and joint training of left and right hand sequences were conducted on the original data according to the mirror symmetric characteristics of human left and right hands.Finally,the generation of fingering was realized by combining the bidirectional long short-term memory network-conditional random field(BiLSTM-CRF)model.Experimental results show that the proposed algorithm is considerably better than commonly used statistical and deep learning methods,which confirms the rationality and effectiveness of the proposed model.
作者
李锵
吴正彪
关欣
LI Qiang;WU Zhengbiao;GUAN Xin(School of Microelectronics,Tianjin University,Tianjin 300072,China)
出处
《智能系统学报》
CSCD
北大核心
2023年第6期1287-1294,共8页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(61872267)
天津市自然科学基金项目(16JCZDJC31100)
天津大学创新基金项目(2021XZC-0024).
关键词
人工智能
音乐
信息检索
长短时记忆
循环神经网络
数据处理
特征提取
时间序列
artificial intelligence
music
information retrieval
long short-term memory
recurrent neural networks
data processing
feature extraction
time series
