一种基于液体状态机的音乐和弦序列识别方法

摘要

A chord sequence recognition algorithm based on Liquid State Machine ( LSM) is presented. Firstly, the music signal is segmented and Pitch Class Profile feature is extracted for every frame. Then, a LSM model is achieved after training. Two kinds of Bizarre Chord, chord appears probability vector and chord transformation matrix, are presented to post-process the chord sequence outputted by LSM. 8 sets of experimental data from neural network model, hidden Markov mode, echo state network model and LSM model show that the LSM gets a good performance, and the post-processing method also effectively improves the recognition accuracy.%　　文中提出一种基于液体状态机的音乐和弦序列识别方法。该方法首先将音乐信号进行切分采样并对每帧提取音级轮廓(PCP)，经训练后得到一个液体状态机模型。方法提出两类奇异矩阵、和弦出现概率向量、和弦变换矩阵，它们可用在和弦序列后处理阶段。在神经网络模型、隐马尔科夫模型、回声状态网络模型、液体状态机模型上进行的初步实验得到8组实验数据。数据表明液体状态机模型对音乐和弦序列具有较好的识别效果，文中提出的后处理算法也能显著提高识别准确率。