Bird Sound Feature Extraction And Phoneme Classification Research

With the increasing influence of human activities, the air is polluted, forest is reducing, land is easy to wash away and the biodiversity is decreasing. The loss of biological populations indicates the changes of ecological environment. As an important group of wild animals, the birds are also affected by the environmental change.12percent of bird species have endangered. At present, there are over9700kinds of birds all over the world, and1244species (944subspecies) of birds in China. The birds are rich in species numbers, that it can be used as marker beacon of the ecological environment. So, the monitoring of bird species changes may indirectly reflect the environment changes of this region.The traditional identification method of bird monitoring always relies on the ornithologists, who hear and note the bird song artificially. But with the development of sensor and computer technology, the bird monitoring has begun to transform to automatic monitoring and recognition. This paper elementarily realizes the extraction and identification of bird song, based on the theoretical analysis of bird song identification. And the main contributions are as follows:1) The implementation method of bird song pre-treatment processes have been presented, such as pre-emphasis, sub-frame and adding window. And the suitable parameters of bird song characteristics have been chose for programming implementation in the Matlab environment;2) According to the comparative analysis of traditional Wiener filtering algorithm and wavelet analysis, this paper chooses the waveform, signal to noise ratio and root mean square error as the evaluation indexes, which has been used to verify that the wavelet analysis is more suit for bird song than wiener filtering;3) A new element segmentation method based on time-frequency characteristics is put forward, using the short-time zero crossing rate and short-term energy as a threshold variable to improve the accuracy of element segmentation;4) The traditional signal extraction method of bird song characteristics just only extracted features from the time domain or frequency domain. So this paper puts forward a new comprehensive extraction method to replace the traditional way. This method improves the differences and representative of the identifying characteristics, and increase the effectiveness and accuracy of classification;5) In order to fit the earlier extracted feature vector of bird song, this paper chooses the organizing map neural network to construct a neural network structure, which has352input neurons,1hidden layer and10output neurons used to realize the automatic classification of bird song elements;6) A bird song processing and analysis software has been designed and developed based on the Matlab graphical guide, which has friendly interface and simple operation. Moreover, it has realized parts of the arithmetic functions such as reading the song file, de-noising, feature extraction, element segmentation and classification. And it provides an operating platform for the experimental study.