Abstract: There are many kinds of sounds in the eco-environment and it is significant to recognize them quickly and accurately，but the recognition difficulty is greatly increased because of the background noise.Therefore，this paper presents a two-layer recognition technology which has good anti-noise ability and high recognition performance.Two kinds of features which are new improved MFCC and fluctuation pattern are selected as feature sets of the eco-environmental sound.The new MFCC parameters are used to construct single Gaussian distribution model and then the Kullback-Leibler distance between the unknown signal’s Gaussian and the sample signal’s Gaussian is measured and subsequently Euclidean distance between the sample signal’s fluctuation pattern and the unknown signal’s is calculated.The calculated Kullback-Leibler distance and the Euclidean distance can be used to achieve a two-layer system for environmental sound recognition.Experimental results show that the two-layer recognition technology can maintain a high recognition rate even under the influence of noise.

Key words: eco-environment, sound recognition, improved Mel-scaled Cepstrum Coefficients（MFCC）, fluctuation pattern, Kullback-Leibler distance

摘要： 对生态环境中各种不同的声音进行快速准确的识别有重要的现实意义，但是因其具有较高背景噪声加大了识别的难度。提出一种具有良好抗噪能力和较高识别性能的两层音频识别技术。选择经过改进的新型的MFCC参数以及波动模型作为生态环境声音的特征集合。利用这种新型的MFCC系数构造音频信号的高斯分布模型，并且计算未知音频信号与样本音频信号的高斯分布模型之间的Kullback-Leibler距离，随后计算它们的波动模型之间的欧几里德距离。根据计算出的Kullback-Leibler距离和欧几里德距离实现两层音频识别系统。实验结果表明两层音频识别技术即使在噪声的影响下也能保持较高的识别率。

关键词: 生态环境, 声音识别, 改进的Mel频率倒谱参数, 波动模型, Kullback-Leibler距离