Abstract: A recognition method for natural sounds based on Gaussian Mixture Model（GMM） distribution is proposed.Mel-Frequency Cepstral Coefficients（MFCCs） are used to analyze natural sounds for their feature extraction.The expectation maximization algorithm is used to learn a Gaussian mixture model distribution of MFCCs for the set of audio feature vectors that describe each sound.Minimum classification error criterion and vote rule are used to yield higher recognition accuracy for natural sounds.Experimentally，compared with K-Nearest Neighbor（KNN） method，GMM is able to achieve a higher accuracy rate for discriminating 36 classes of natural sounds.The classified accuracy rate of GMM reaches to 95.83%.

Key words: Mel-frequency cepstral coefficients, Gaussian mixture model, natural sounds recognition, vote rule

摘要： 提出了一种基于高斯混合模型（GMM）的自然环境声音的识别方法。提取Mel频率倒谱系数（MFCCs）来分析声音信号;对于每种声音使用期望最大化算法基于MFCC特征集建立高斯混合模型;使用最小错误率判决规则和投票裁决的方法进行识别。使用GMM对36种自然环境的声音进行识别的正确率可达95.83%，且识别效果优于K最近邻（KNN）。

关键词: Mel频率倒谱系数, 高斯混合模型, 自然环境声音的识别, 投票裁决