摘要
针对音符识别准确率不高的问题,提出一种基于人工神经网络的音符识别方法。为提高识别准确率,首先采用归一化和汉明窗方法对音频信号进行预处理,然后采用CQT和MFCC分别提取频域特征和倒频域特征;利用BP神经网络和Softmax回归模型,提出Softmax回归结合BP神经网络音符识别模型,并构建音符识别分类器;最后通过MATLAB R2016a作为仿真软件,在自构音符库的基础上,对音符进行识别。结果表明,在CQT和MFCC共同提取特征和不同样本数量下,本研究构建的音符识别器的识别率都高于93%,且与其他参考文献的识别率相比,本研究算法也具有明显优势。由此说明,本研究构建的音符特征提取与识别方案可行。
Aiming at the low accuracy of note recognition, a note recognition method based on artificial neural network is proposed. In order to improve the recognition accuracy, the audio signal is preprocessed by normalization and Hamming window methods, and then the frequency domain features and inverse frequency domain features are extracted by CQT and MFCC respectively;Using BP neural network and softmax regression model, a note recognition model combining softmax regression and BP neural network is proposed, and a note recognition classifier is constructed;Finally, matlabr r2016 a is used as the simulation software to recognize the notes on the basis of the self constructed sound library. The results show that under the joint extraction of features by CQT and MFCC and different sample numbers, the recognition rate of the note recognizer constructed in this study is higher than 93%, and compared with the recognition rate of other references, this research algorithm also has obvious advantages. This shows that the note feature extraction and recognition scheme constructed in this study is feasible.
作者
侯清睿
安冬
HOU Qingrui;AN Dong(Shaanxi Technical College f Finance and Economics,Xianyang,Shaanxi 712000,China)
出处
《自动化与仪器仪表》
2022年第1期53-58,共6页
Automation & Instrumentation
基金
陕西省自然科学类基金项目:BIM技术在建筑工程施工现场布置中的应用研究(ZK19-30)。