计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (11): 155-160.DOI: 10.3778/j.issn.1002-8331.1611-0044

• 模式识别与人工智能 • 上一篇    下一篇

范数正则化解相关集成学习基音频率检测

张小恒1,2,李勇明2,朱  斌2   

  1. 1.重庆广播电视大学,重庆 400052
    2.重庆大学 信号与信息处理研究所,重庆 400030
  • 出版日期:2017-06-01 发布日期:2017-06-13

L2 decorrelated neural network ensembles based pitch tracking

ZHANG Xiaoheng1,2, LI Yongming2, ZHU Bin2   

  1. 1.Chongqing Radio & TV University, Chongqing 400052, China
    2.Research Institute of Signal and Information Processing, Chongqing University, Chongqing 400030, China
  • Online:2017-06-01 Published:2017-06-13

摘要: 低信噪比环境下的基音频率检测极其重要且富有挑战性,至今未得到很好的解决。基于此,首先构造了基于PEFAC的频域空间检测模型,将基音频率作为特征进行提取,然后提出范数正则化的解相关集成学习神经网络模型(L2-DNNE)对其进行训练,利用负相关学习机制(NCL)和模型复杂度约束项提高集成学习模型的泛化能力,从而获取基音频率的最优值,且在测试精度和时间代价上取得了较好的平衡。将该算法与相关有代表性的算法进行比较。比较结果表明,该算法在不同类型不同程度的噪声环境下,能显著提升检测识别率,尤其在低信噪比下有更显著优势。

关键词: 低信噪比环境, 基音频率, 范数正则化的解相关集成学习神经网络模型(L2-DNNE)

Abstract: Fundamental frequency determination in low SNR noise environment is a challenging job, and has not been got solved well so far. Based on this, in this paper, firstly it builds a PEFAC based frequency-domain detection model, and then extracts the characteristic values of fundamental frequency. After that, a L2-DNNE based regression model is proposed, which can ensure the generation ability based on the NCL and model complexity adjustment, and beneficial to searching of the optimum, moreover the algorithm can obtain a balance on test accuracy and time cost. At last, it compares the performance of the algorithm with that of other representative algorithm. The experimental results show that it performs well especially in high levels of additive noise.

Key words: very low Signal Noise Ratio(SNR) environment, fundamental frequency, L2 Decorrelated Neural Network Ensembles(L2-DNNE)