Specific Environmental Sounds Recognition Using Time-frequency Texture Features

Ecological environment is closely related to human life, environmental sounds recognition can learn the life activities of some specific species, as well as do some analysis to the relevant environment. Audio files are easy to record and store, so it makes environmental sounds recognition easy to realize and develop fast recently. However, most of the environmental sounds are non-stationary, and the features chosen in most of the existing environmental sounds recognition methods are based in single time domain or frequency domain, which makes them fail to handle with non-stationary signals. At the same time, the non-stationary noises in the real environment also increases the difficulty. Therefore, the paper studies a specific environmental sounds recognition method using time-frequency texture features and random forest in the non-stationary noise environment, the main work of which includes the followings:(1) Front-end audio enhancement processing. On account of the non-stationary noises in the real environment, this paper proposes an audio enhancement algorithm for the highly non-stationary environmental noises by combining the classical short-time spectrum estimation algorithm with a dynamic noise power spectrum estimation algorithm based on two-way searching, and use this algorithm to do the front-end audio enhancement of specific environmental sounds.(2) Time-frequency texture features extraction. After analyzing to these specific environmental sounds signals, this paper firstly applies a double threshold endpoint detection algorithm based on short-time energy to remove the silence, to solve the problems that the silence contains in the enhanced time-domain audio signals needs more space and increase the computation. And then turns it into its time-frequency spectrum form. At the last, by comparing GLCM, which is the common used method to extract time-frequency texture features, with SDH, which is a modified method based on GLCM, on feathers storage and secondary calculation, chooses to use SDH to make time-frequency texture features extraction, in view of the big space and computation in GLCM. For every specific environmental sound signal, calculate 5 texture features in one of the 4 different relation of position, and achieve a time-frequency texture feature vector whose dimension is 1 and length is 20.(3) Recognition and classification. According to deep researching to the common used single classifiers, especially the decision tree, this paper uses a assemble classifier based on decision tree, namely random forest to make recognition and classification to specific environmental sounds, to solve the problems of unbalance performance to different kinds of data and long response time et al. The training and test of random forest are both based on decision tree, this makes it inherit the strong points of decision tree, and overcome the weak points of decision tree effectively.Three contrast experiments are set to in total of 51 kinds of specific environmental sounds of weather, birds, mammals and insects, the results of which show that the algorithm proposed in this paper can make recognition of the specific environmental sounds with the non-stationary noises fast and effectively.