基于改进MFCC的鸟鸣声识别方法研究

Research of Birdsong Recognition Method Based on Improved MFCC

鸟鸣声识别的关键就在于对鸣声信号进行合理的特征值提取。鸟鸣声信号具有非平稳性,传统的梅尔倒谱系数(MFCC)仅能反映鸣声信号的静态特性,并且算法中直接运用FFT处理非平稳信号存在局限性。本文提出了一种基于经验模态分解(EMD)改进的MFCC算法,通过对鸟鸣声信号进行经验模态分解,得到一系列固有模态函数分量后再进行FFT计算,频域合成后通过Mel滤波器,取其对数能量进行DCT变换,然后对结果作差分得到改进的MFCC参数,再采用高斯混合模型(GMM)进行鸟鸣声的识别。实验结果表明,改进的MFCC识别率达到70.09%,与未改进的MFCC识别率相比提高了3.42%。

To choose a proper feature extraction method is the key of birdsong recognition.The signals of birdsong are non-stationary.The conventional Mel-Frequency Cepstral Coefficients(MFCC)can only reflect signals’static features and has a certain limitation to use FFT to process the signals directly.In this paper,an improved MFCC algorithm is raised on the basis of Empirical Mode Decomposition(EMD).FFT is used after divided signals of birdsonginto intrinsic mode functions with EMD.The next steps are frequency synthesis,Mel filtering,logarithm and DCT.In the end,the improved MFCC parameters are obtained by doing differential calculation of the DCT results.The birdsongs can be recognized through the Gaussian Mixture Model(GMM).The results show that the recognition rate of the improved MFCC is 70.09%,and the recognition rate is increased by 3.42%compared with the conventional MFCC.