Abstract: Mel frequency cepstrum coefficients, motivated by human auditory, are widely used for sound recognition, while in the ecological environment the presence of the noise causes its recognition rate to drop quickly. This paper presents a method for the classification of natural sounds in noise environment. Multi-band spectral subtraction method is proposed to reduce the noise for extracting MFCC, which effectively weakens audio noise disturbance. In addition, it introduces SVM classification model to classify the sound with noise owing to its better noise immunity and robustness. Experimental results show that this method can be used to classify ecological environmental sounds accurately.

Key words: multi-band spectral subtraction, eco-environment sounds classification, Mel frequency cepstrum coefficients, Support Vector Machine（SVM）

摘要： 基于人类听觉特性的Mel频率倒谱系数广泛用于声音识别，然而在生态环境中噪声的出现导致其识别率剧减。提出一种在噪声背景下生态环境声音分类方法。利用非线性多频带谱减法对声音功率谱进行去噪处理并提取改进Mel频率倒谱系数，有效削弱不同频率段噪声功率谱干扰。利用支持向量机良好的鲁棒性和抗噪能力对含有噪声的生态环境声音进行分类。实验表明该方法能有效在噪声背景下对生态环境声音准确分类。

关键词: 多频带减谱法, 生态环境声音分类, Mel频率倒谱系数, 支持向量机