基于改进深度学习算法的乐器音色识别

Musical Instrument Timbre Recognition Based on Improved Deep Learning Algorithm

为保持乐器音色时间序列基础上实现乐器音色的准确识别,提出基于改进深度学习算法的乐器音色识别新方法。该方法首先采用基于梅尔滤波器组能量对数和梅尔频率倒谱系数的一维卷积神经网络提取乐器音色特征;其次,将乐器音色特征输入基于长短期记忆和深度神经网络的乐器音色分类器进行乐器音色识别;最后,对来自5个乐器声音数据库的乐器音色进行了仿真测试。对于5个数据库中乐器音色的识别测试结果表明,该乐器音色识别比基于卷积神经网络的乐器音色识别方法、基于卷积神经网络与深度置信网络的乐器音色识别方法分别提高了2.49%和2.02%。

To achieve accurate identification of musical instrument timbres while preserving the temporal characteristics of their sound,a new approach based on an improved deep learning algorithm is proposed.This method commences by utilizing a one-dimensional convolutional neural network to extract features of musical instrument timbres,which leverages the combination of Mel-scaled filter bank log energies and Mel-Frequency Cepstral Coefficients.Then input the musical instrument timbre features into the musical instrument timbre classifier based on long short-term memory and deep neural network for musical instrument timbre recognition.The results of the recognition tests conducted on the timbres of instruments from these five databases reveal that the proposed approach outperforms both traditional convolutional neural network-based timbre recognition methods and hybrid convolutional neural network deep belief network approaches.Specifically,it achieves an improvement of 2.49%over convolutional neural network methods and 2.02%over hybrid convolutional neural network deep belief network methods,demonstrating its effectiveness in accurately identifying musical instrument timbres while preserving the inherent temporal dynamics of their sounds.