基于MFCC和GMM的昆虫声音自动识别

摘要

Insects produce various sounds when they are moving, feeding or calling. These sounds exhibit intraspecies similarity and interspecies differences, thus they can be used to discriminate species identities of insects. Automatic detection of insect species through sounds produced by the insects would be very meaningful in giving farm workers or forestry workers a convenient way to recognize insects. In this study we employed the sound parameterization techniques that are frequently used in the field of human speech recognition. Mel-frequency cepstrum coefficients (MFCCs) were extracted from the sound samples after preprocessing, and Gaussian mixture model (GMM) was trained with these MFCC features. Finally, the unknown insect sound samples were classified by the GMM. The proposed method was evaluated in a database with acoustic samples of 58 different insect sounds. The method performed well in terms of both recognition rate and time performance. The average recognition accuracy was as high as 98. 95%. The test results proved that sound parameterization techniques based on MFCC and GMM could be used to recognize insect species efficiently.%昆虫的运动、取食、鸣叫都会发出声音,这些声音存在种内相似性和种间差异性,因此可用来识别昆虫的种类.基于昆虫声音的昆虫种类自动检测技术对协助农业和林业从业人员方便地识别昆虫种类非常有意义.本研究采用了语音识别领域里的声音参数化技术来实现昆虫的声音自动鉴别.声音样本经预处理后,提取梅尔倒谱系数(Mel-frequency cepstrum coefficient,MFCC)作为特征,并用这些样本提取的MFCC特征集训练混合高斯模型(Gaussian mixture model,GMM).最后用训练所得到的GMM对未知类别的昆虫声音样本进行分类.该方法在包含58种昆虫声音的样本库中进行了评估,取得了较高的识别正确率(平均精度为98.95％)和较理想的时间性能.该测试结果证明了基于MFCC和GMM的语音参数化技术可以用来有效地识别昆虫种类.