隐式特征和循环神经网络的多声部音乐生成系统

PDF下载HTML

Title:

A polyphony music generation system based on latent features and a recurrent neural network

作者:

苗北辰¹, 郭为安², 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 中德学院, 上海 201804

Author(s):

MIAO Beichen¹, GUO Weian², WANG Lei¹

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

关键词:

音乐生成; 隐式特征提取; 循环神经网络; 栈式自编码器; 多声部音乐; 序列预测; 长短期记忆循环神经网络; 生成模型

Keywords:

music generation; latent feature extraction; recurrent neural network; stacked autoencoder; polyphony music; sequence prediction; long short-term memory; generation model

分类号:

TP393.04

DOI:

10.11992/tis.201804009

摘要:

音乐生成是一种使用算法来生成音乐序列的研究。本文针对音乐样本特征提取以及自动作曲问题提出了一种基于音乐隐式特征和循环神经网络（recurrent neural network， RNN）的多声部音乐生成算法。该方法通过使用栈式自编码器对多声部音乐序列每个时间步的音符隐式特征进行提取，结合长短期记忆循环神经网络（long short-term memory， LSTM），以序列预测的方式搭建了基于隐式特征的音乐生成模型。仿真结果表明，该音乐生成算法在使用相同风格的音乐数据训练后，得到的模型可以生成旋律与和弦匹配较好的多声部音乐数据。

Abstract:

Music generation is a research area that uses algorithms to generate sequences with characteristics of music. Focusing on the problem of feature extraction from music samples and automatic music compositions, this paper proposes a polyphony music generation algorithm based on musical latent features and a recurrent neural network (RNN). The proposed algorithm uses a stacked autoencoder to extract latent features from of music sequence notes at each time step; the algorithm then uses long-short term memory RNNs to build a music generation system in the form of sequence prediction. The simulation results show that this algorithm can generate polyphony music with better melody and chord matching.