基于优化神经网络的音乐分类模型研究

黄璟

摘  要： 为高效、精准地判断音乐风格归属，帮助用户快速获取偏好音乐风格，基于群智优化神经网络构建音乐风格分类模型。提取音乐样本的音质、节奏、旋律特征作为模型训练样本输入BP神经网络，通过初始化、隐含层及输出层计算、权值计算等步骤完成神经网络模型训练。采用粒子群算法确定神经网络的最优权值与阈值，粒子群算法首先编码神经网络权值与阈值，其次计算粒子适应度值，更新粒子速度和位置，符合终止条件时输出神经网络的权值与阈值优化结果，并据此构建基于群智优化神经网络的音乐风格分类模型。模型测试结果表明，所提模型在正确区分不同音乐风格的同时，展示了音质、节奏、旋律等特征。

关键词： 音乐分类; 分类模型; 特征提取; 模型训练; 最优权值确定; 模型构建

中图分类号： TN911.1?34; TP181                   文献标识码： A                    文章编号： 1004?373X（2020）03?0096?04

Research on music classification model based on optimal neural network

HUANG Jing

（Music Department， Putian University， Putian 351100， China）

Abstract： In order to efficiently and accurately judge the type of music style and help users quickly obtain their preference for music style， a music style classification model was built based on swarm intelligence optimization neural network. The acoustic quality， rhythm and melody characteristics of music samples are extracted as model training samples and input into BP neural network. The neural network model training is completed after the steps of initialization， hidden layer and output layer calculation， weight calculation and so on. The particle swarm algorithm is adopted to determine the optimal neural network weight and threshold， and then code the neural network weight and threshold， and calculate particle fitness value to update the particle velocity and position. When they conform to with the termination conditions， the optimization results of the neural network weight and threshold are output. On the basis of the results， the music style classification model based on swarm intelligence optimization neural network is constructed. The model testing results show that the model can correctly distinguish different music styles， and exhibits the characteristics of sound quality， rhythm and melody.

Keywords： music classification; classification model; feature extraction; model training; optimal weight determination; model building

0  引  言

當代互联网技术高速更新，移动终端设备使用人群规模迅速增加，流行音乐传播方式从电台和唱片形式逐渐过渡到网络电台在线收听和互联网下载等形式。相关调查显示，65%用户难以精准获取期待收听的歌曲或不明确自身音乐风格喜好。网络电台通常依照用户对音乐风格的喜好推荐歌曲，用户喜好的音乐类型可能包含多种音乐风格，导致待分类音乐规模较大，音乐分类难度大。因此，采用一种智能方法实现音乐风格分类十分重要[1]。流行音乐传统分类包括流行、乡村、爵士、摇滚、节奏布鲁斯、新世纪等六大类别，每一大类划分成不同小类，部分小类还可以进一步划分[2]。……