A Feature Extraction Method Fused with Inter-class Standard Deviation in Acoustic Scene Classification

Abstract

The invention provides a feature extraction method integrating between-class standard deviation in sound scene classification. The feature extraction method comprises the following steps: S1, carrying out feature extraction based on a traditional mode, namely, calculating a frequency spectrum diagram of original audio, and carrying out downsampling based on a traditional filter, thus obtaining a feature frequency spectrum diagram P1 after downsampling; S2, carrying out feature extraction based on between-class standard deviation, namely, calculating a frequency spectrum diagram of original audio, and carrying out downsampling based on a between-class frequency domain standard deviation filter, thus obtaining a between-class standard deviation feature frequency spectrum diagram P2 after downsampling; and S3, carrying out feature fusion based on the between-class standard deviation, namely, splicing the feature frequency spectrum diagram P1 in S1 and the feature frequency spectrum diagram P2 in S2, wherein the splicing result is taken as input for a sound scene classification model. With the technical scheme for improving the sound scene classification accuracy rate, the problem that the existing sound scene resolution is not high is solved; the features are extracted through the between-class standard deviation for the first time and are fused with other features, and therefore, the identification property of the system is improved. The system provided by the invention is simple in structure and convenient to implement.

Description

technical field [0001] The invention relates to the field of sound signal analysis, in particular to a feature extraction method for integrating standard deviations between classes in sound scene classification. Background technique [0002] In recent years, in the field of audio research, under the attention of many scholars, the task of speech recognition has made great progress. However, non-speech sounds such as environmental sounds also contain important information, so their analysis and understanding are also important. of equal importance. The concept of Acoustic scene classification (ASC) is to identify the environment in which the voice clip was recorded by analyzing the voice clip, and assign the corresponding environmental semantic label to the audio. Such as parks, subways, offices, etc. The main research goal of ASC is to enable computers to understand the surrounding environment by analyzing sound like the human auditory system. It is a research direction re...

AI Technical Summary

Problems solved by technology

The common extraction feature in audio is the mel spectrum. The mel spectrum is a spectrogram extracted based on the resolution of the human ear's perception of frequency, and the resolution of the acoustic scene of each frequency component may not be completely consistent with the perceptual resolution. Only a single feature is used. Spectrum is used as CNN feature input, and there is a problem of insufficient feature expression, which will affect the recognition rate of acoustic scene classification

Images